CN108875487A - Pedestrian is identified the training of network again and is identified again based on its pedestrian - Google Patents
Pedestrian is identified the training of network again and is identified again based on its pedestrian Download PDFInfo
- Publication number
- CN108875487A CN108875487A CN201710906719.5A CN201710906719A CN108875487A CN 108875487 A CN108875487 A CN 108875487A CN 201710906719 A CN201710906719 A CN 201710906719A CN 108875487 A CN108875487 A CN 108875487A
- Authority
- CN
- China
- Prior art keywords
- network
- pedestrian
- loss
- training
- negative sample
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides method, apparatus, system and storage medium that pedestrian is identified the training of network and identified based on its pedestrian again again, the pedestrian identifies that the training method of network includes again:Pre-training is carried out to baseline network using Classification Loss;And joint classification loss and five-tuple loss carry out tuning to pre-trained baseline network and identify network again to obtain pedestrian.Pedestrian according to an embodiment of the present invention identifies that training method, device, system and the loss of storage medium joint classification of network and range loss are trained again, can accelerate training process and improve precision;In addition, using five-tuple method in range loss link, compared to traditional triple, triple and four-tuple method are improved, can significantly shorten the training time, and further increase precision.
Description
Technical field
The present invention relates to pedestrian's weight identification technology fields, relate more specifically to training and base that a kind of pedestrian identifies network again
In its pedestrian recognition methods, device, system and storage medium again.
Background technique
Pedestrian identify again also referred to as pedestrian identify again, be judged using computer vision technique be in image or video sequence
It is no that there are the technologies of specific pedestrian.A monitoring pedestrian image is given, the pedestrian image under striding equipment is retrieved.It is intended to make up mesh
The vision of the camera of preceding fixation is limited to, and can be combined with pedestrian detection/pedestrian tracking technology, can be widely applied to intelligent view
The fields such as frequency monitoring, intelligent security.
Existing pedestrian again recognition methods according to training thought be segmented into two kinds of ways:The first is that each pedestrian is made
For a classification, pedestrian is identified again and is converted into image classification problem;Second is the feature for extracting every pedestrian's picture, is calculated
The distance of two kinds of picture features is schemed between the distance of picture feature and the different pedestrians of maximization between the same person by minimizing
The distance of piece feature come train one extraction feature network model, current method include triple, improve triple and
Four-tuple.
However, the model based on Classification Loss training is extremely difficult to a high level in precision aspect, and based on away from
Although the model precision from loss training would generally be better than the former, net training time is but very long.
Summary of the invention
In view of the above-mentioned problems, the invention proposes a kind of schemes of training for identifying network again about pedestrian, in conjunction with two
The advantages of kind method, to accelerate training process and improves precision by joint classification loss and range loss.It is briefly described below
The scheme of the training proposed by the present invention for identifying network again about pedestrian, more details will be embodied in subsequent combination attached drawing
It is described in mode.
According to an aspect of the present invention, the training method that a kind of pedestrian identifies network again is provided, the training method includes:
Pre-training is carried out to baseline network using Classification Loss;And joint classification loss and five-tuple loss are to pre-trained benchmark
Network carries out tuning and identifies network again to obtain pedestrian.
In one embodiment of the invention, described to include to baseline network progress pre-training using Classification Loss:By sample
This picture is input to the baseline network;Predicted vector and the sample by the baseline network for samples pictures output
The label vector of this picture is compared to obtain Classification Loss;The ginseng of the baseline network is adjusted based on the Classification Loss
Number;And above-mentioned steps are repeated, until classification accuracy and Classification Loss no longer change substantially.
In one embodiment of the invention, the baseline network is residual error network.
In one embodiment of the invention, before the samples pictures are input to the baseline network, to described
Samples pictures implement pretreatment operation.
In one embodiment of the invention, the joint classification loss and five-tuple loss are to pre-trained reference net
Network carries out tuning:By pre-provisioning request and five samples pictures for sequentially inputting five-tuple;It is directed to based on the baseline network
The predicted vector of every samples pictures output calculates Classification Loss;Five sample graphs are directed to based on the baseline network
The feature vector of piece output calculates five-tuple loss;And based on Classification Loss calculated and five-tuple calculated loss
Loss of the final loss to identify network again as the pedestrian.
In one embodiment of the invention, the Classification Loss calculated is that the classification of five samples pictures is damaged
The average value of mistake.
In one embodiment of the invention, the five-tuple loss is defined as:
lqt=d (positive sample 1, positive sample 2)-d (negative sample 1, negative sample 21)+d (negative sample 21, negative sample 22)-d is (negative
Sample 1, positive sample 2)+a
Wherein, lqtFor five-tuple loss;Positive sample 1, positive sample 2, negative sample 1, negative sample 21 and negative sample 22 are institute
Five samples pictures are stated, and positive sample 1 and positive sample 2 are two different pictures of the first pedestrian, negative sample 1 is the second pedestrian's
Picture, negative sample 21 and negative sample 22 are two different pictures of third pedestrian;D be two pictures feature vector between away from
From;A is the constant parameter being arranged according to demand.
In one embodiment of the invention, the final loss by the Classification Loss calculated and described is counted
The weighted sum of the five-tuple loss of calculation.
According to a further aspect of the invention, the training device that a kind of pedestrian identifies network again is provided, the experienced device includes:
Pre-training module, for carrying out pre-training to baseline network using Classification Loss;And tuning module, it is lost for joint classification
Tuning is carried out to pre-trained baseline network with five-tuple loss and identifies network again to obtain pedestrian.
In one embodiment of the invention, the pre-training module further wraps the pre-training of the baseline network
It includes:Samples pictures are input to the baseline network;The baseline network is directed to the predicted vector of samples pictures output
It is compared with the label vector of the samples pictures to obtain Classification Loss;The reference net is adjusted based on the Classification Loss
The parameter of network;And aforesaid operations are repeated, until classification accuracy and Classification Loss no longer change substantially.
In one embodiment of the invention, the baseline network is residual error network.
In one embodiment of the invention, the pre-training module is also used to:The samples pictures are being input to institute
Before stating baseline network, pretreatment operation is implemented to the samples pictures.
In one embodiment of the invention, the tuning module includes to the tuning of pre-trained baseline network:It presses
Pre-provisioning request and five samples pictures for sequentially inputting five-tuple;It is defeated for every samples pictures based on the baseline network
Predicted vector out calculates Classification Loss;Based on feature vector of the baseline network for five samples pictures output
Calculate five-tuple loss;And based on the final loss of Classification Loss calculated and five-tuple costing bio disturbance calculated using as
The pedestrian identifies the loss of network again.
In one embodiment of the invention, the Classification Loss calculated is that the classification of five samples pictures is damaged
The average value of mistake.
In one embodiment of the invention, the five-tuple loss is defined as:
lqt=d (positive sample 1, positive sample 2)-d (negative sample 1, negative sample 21)+d (negative sample 21, negative sample 22)-d is (negative
Sample 1, positive sample 2)+a
Wherein, lqtFor five-tuple loss;Positive sample 1, positive sample 2, negative sample 1, negative sample 21 and negative sample 22 are institute
Five samples pictures are stated, and positive sample 1 and positive sample 2 are two different pictures of the first pedestrian, negative sample 1 is the second pedestrian's
Picture, negative sample 21 and negative sample 22 are two different pictures of third pedestrian;D be two pictures feature vector between away from
From;A is the constant parameter being arranged according to demand.
In one embodiment of the invention, the final loss by the Classification Loss calculated and described is counted
The weighted sum of the five-tuple loss of calculation.
According to a further aspect of the invention, provide a kind of pedestrian's recognition methods again, the pedestrian again recognition methods using upper
It states described in any item pedestrians and identifies that pedestrian made of the training method training of network identifies that network carries out pedestrian and identifies again again again.
According to a further aspect of the invention, a kind of pedestrian's weight identification device is provided, pedestrian's weight identification device is for real
Apply above-mentioned pedestrian recognition methods again.
Another aspect according to the present invention provides a kind of computing system, and the system comprises storage device and processor, institutes
The computer program for being stored on storage device and being run by the processor is stated, the computer program is transported by the processor
Pedestrian described in any of the above embodiments is executed when row to identify the training method of network again or execute above-mentioned pedestrian's recognition methods again.
According to a further aspect of the present invention, a kind of storage medium is provided, is stored with computer program on the storage medium,
The computer program executes pedestrian described in any of the above embodiments at runtime and identifies the training method of network again or execute above-mentioned
Pedestrian's recognition methods again.
Pedestrian according to an embodiment of the present invention identifies training method, device, system and the storage medium joint point of network again
Class loss and range loss are trained, and can be accelerated training process and be improved precision;In addition, being used in range loss link
Five-tuple method compared to traditional triple, improves triple and four-tuple method, can significantly shorten the training time, and
And further increase precision.
Detailed description of the invention
The embodiment of the present invention is described in more detail in conjunction with the accompanying drawings, the above and other purposes of the present invention,
Feature and advantage will be apparent.Attached drawing is used to provide to further understand the embodiment of the present invention, and constitutes explanation
A part of book, is used to explain the present invention together with the embodiment of the present invention, is not construed as limiting the invention.In the accompanying drawings,
Identical reference label typically represents same parts or step.
Fig. 1 show for realizing pedestrian according to an embodiment of the present invention identify again the training method of network, device, system and
The schematic block diagram of the exemplary electronic device of storage medium;
Fig. 2 shows the schematic flow charts that pedestrian according to an embodiment of the present invention identifies the training method of network again;
Fig. 3 shows the schematic diagram of baseline network pre-training according to an embodiment of the present invention;
Fig. 4 shows the schematic diagram of tuning after baseline network pre-training according to an embodiment of the present invention;
Fig. 5 shows the schematic block diagram that pedestrian according to an embodiment of the present invention identifies the training device of network again;And
Fig. 6 shows the schematic block diagram that pedestrian according to an embodiment of the present invention identifies the training system of network again.
Specific embodiment
In order to enable the object, technical solutions and advantages of the present invention become apparent, root is described in detail below with reference to accompanying drawings
According to example embodiments of the present invention.Obviously, described embodiment is only a part of the embodiments of the present invention, rather than this hair
Bright whole embodiments, it should be appreciated that the present invention is not limited by example embodiment described herein.Based on described in the present invention
The embodiment of the present invention, those skilled in the art's obtained all other embodiment in the case where not making the creative labor
It should all fall under the scope of the present invention.
Firstly, identifying the training method of network, dress again to describe the pedestrian for realizing the embodiment of the present invention referring to Fig.1
It sets, the exemplary electronic device 100 of system and storage medium.
As shown in Figure 1, electronic equipment 100 include one or more processors 102, it is one or more storage device 104, defeated
Enter device 106, output device 108 and image collecting device 110, these components pass through bus system 112 and/or other forms
Bindiny mechanism's (not shown) interconnection.It should be noted that the component and structure of electronic equipment 100 shown in FIG. 1 are only exemplary,
And not restrictive, as needed, the electronic equipment also can have other assemblies and structure.
The processor 102 can be central processing unit (CPU) or have data-handling capacity and/or instruction execution
The processing unit of the other forms of ability, and the other components that can control in the electronic equipment 100 are desired to execute
Function.
The storage device 104 may include one or more computer program products, and the computer program product can
To include various forms of computer readable storage mediums, such as volatile memory and/or nonvolatile memory.It is described easy
The property lost memory for example may include random access memory (RAM) and/or cache memory (cache) etc..It is described non-
Volatile memory for example may include read-only memory (ROM), hard disk, flash memory etc..In the computer readable storage medium
On can store one or more computer program instructions, processor 102 can run described program instruction, to realize hereafter institute
The client functionality (realized by processor) in the embodiment of the present invention stated and/or other desired functions.In the meter
Can also store various application programs and various data in calculation machine readable storage medium storing program for executing, for example, the application program use and/or
The various data etc. generated.
The input unit 106 can be the device that user is used to input instruction, and may include keyboard, mouse, wheat
One or more of gram wind and touch screen etc..
The output device 108 can export various information (such as image or sound) to external (such as user), and
It may include one or more of display, loudspeaker etc..
Described image acquisition device 110 can acquire the desired image of user (such as photo, video etc.), and will be adopted
The image of collection is stored in the storage device 104 for the use of other components.Image collecting device 110 can be camera.
It should be appreciated that image collecting device 110 is only example, electronic equipment 100 can not include image collecting device 110.This
In the case of, it can use other image acquisition device samples pictures or samples pictures, and the picture of acquisition is sent to electricity
Sub- equipment 100.
Illustratively, training method, device, the system of network are identified again for realizing pedestrian according to an embodiment of the present invention
It may be implemented as smart phone, tablet computer etc. with the exemplary electronic device of storage medium.
In the following, reference Fig. 2 is described the training method 200 that pedestrian according to an embodiment of the present invention identifies network again.Such as Fig. 2
Shown, pedestrian identifies that the training method 200 of network may include steps of again:
In step S210, pre-training is carried out to baseline network using Classification Loss.
In one embodiment, pre-training network model can be carried out first with Classification Loss, due to usually passing through tens
Secondary repetitive exercise can make network fast convergence, and reach identical performance, and the method based on range loss at least needs
Ten times of training time is expended, therefore the training time can be greatly shortened using Classification Loss pre-training network model.
In one embodiment, it is known as baseline network using the network model that Classification Loss carries out pre-training, it is subsequent to retouch
The tuning step stated is implemented on the baseline network after pre-training.Illustratively, baseline network can be residual error network,
The for example, residual error network (ResNet50) of large-scale image identification challenge match (ImageNet) pre-training.When baseline network is to be somebody's turn to do
When residual error network, before samples pictures are input to baseline network to be trained, first samples pictures can be pre-processed.
For example, can be 224 × 224 pixels by the size conversion of samples pictures, picture format be BGR channel format, often
A channel needs to cut all images of ImageNet in the average value in the channel, is formulated as:
The channel B -104.00698793 of new channel B=original
The channel G -116.66876762 in the new channel G=original
The channel R -122.67891434 in the new channel R=original
Pretreated as above journey is merely exemplary, and is not required.It in other examples, can also be using other certainly
The convolutional network of definition or other suitable networks are as baseline network, correspondingly, samples pictures can be input to the base
Implement other suitable preprocessing process before pseudo-crystalline lattice.
In one embodiment, carrying out pre-training to baseline network using Classification Loss in step S210 can further wrap
It includes:Samples pictures are input to baseline network;Predicted vector and the sample by baseline network for samples pictures output
The label vector of this picture is compared to obtain Classification Loss;The ginseng of the baseline network is adjusted based on the Classification Loss
Number;And above-mentioned steps are repeated, until classification accuracy and Classification Loss no longer change substantially.
Specifically, the structure and pre-training process of said reference network can be further understood in conjunction with Fig. 3.As shown in figure 3,
Input picture (such as by pretreated samples pictures) can be input to baseline network and (be shown in Figure 3 for residual error net
Network ResNet50), after normalizing classification layer (Softmax), baseline network can be for every samples pictures output one
A predicted vector, the value of i-th of element of the predicted vector indicate this picture be i-th of people probability (i=1,2,
3 ... ..., N, wherein N be natural number), so the element of this vector and be 1.
It then, can be by label vector (label marked, for example, artificial mark of the predicted vector and the samples pictures
Label) compare to obtain Classification Loss.Since label vector is only hot (one-hot) vector, i.e. only one element is
1, other elements are 0, this 1 with regard to representing is which individual, i.e. id information.Classification Loss is the prediction of baseline network output
Difference between vector and label vector (for example, by using entropy loss is intersected).It then, can be by Classification Loss backpropagation Hui Ji
Pseudo-crystalline lattice adjusts the parameter of baseline network.
Forward calculation predicted vector and the reversed network parameter that updates are primary complete iteration, and such iteration is repeated
Classification accuracy and Classification Loss to the last no longer changes then deconditioning substantially.This usual stage only needs tens times
Iterative network can fast convergence.Therefore, the training time can be greatly shortened.
Now referring back to Fig. 2, the training method 200 that pedestrian according to an embodiment of the present invention identifies network again is continued to describe
Subsequent step.
In step S220, joint classification loss and five-tuple loss carry out tuning to pre-trained baseline network to obtain
Pedestrian identifies network again.
In one embodiment, five-tuple is referred to according to five samples for having to the different pedestrians of summation sequential selection three
This picture, concrete condition are as follows:
(1) picture 1:The first picture of pedestrian 1, is named as positive sample 1;
(2) picture 2:The second picture of pedestrian 1, it is different from picture 1, it is named as positive sample 2;
(3) picture 3:The first picture of pedestrian 2, is named as negative sample 1;
(4) picture 4:The first picture of pedestrian 3, is named as negative sample 21;
(5) picture 5:The second picture of pedestrian 3, it is different from picture 4, it is named as negative sample 22.
In one embodiment, five-tuple loss can be defined as formula:
lqt=d (positive sample 1, positive sample 2)-d (negative sample 1, negative sample 21)+d (negative sample 21, negative sample 22)-d is (negative
Sample 1, positive sample 2)+a
Wherein, lqtFor five-tuple loss, positive sample 1, positive sample 2, negative sample 1, negative sample 21 and negative sample 22 are upper
Five samples pictures are stated, positive sample 1 and positive sample 2 are two different pictures of the first pedestrian, and negative sample 1 is the figure of the second pedestrian
Piece, negative sample 21 and negative sample 22 are two different pictures of third pedestrian;A be arranged according to demand constant parameter (such as
May be configured as 2 or other values for being arbitrarily arranged according to actual needs);D is the distance between the feature vector of two pictures, example
Such as, d (positive sample 1, positive sample 2) be pedestrian 1 the first picture and pedestrian 1 the second picture feature vector between away from
The distance between the first picture from the first picture and pedestrian 3 that, d (negative sample 1, negative sample 21) is pedestrian 2, it is all so
Class etc..In one example, above-mentioned d can indicate Euclidean distance.In other examples, can also based on feature vector it
Between other distances come calculate five-tuple loss, COS distance, mahalanobis distance etc..
The distance between feature vector (also referred to as image content feature vector) of different pictures (such as it is two-dimentional it is European away from
From) it can define similarity between different pictures.Above-mentioned samples pictures are input to the baseline network by pre-training, it is complete to connect
Layer Fc (alternatively referred to as characteristic layer, as shown in Figure 3) feature vector for corresponding to every samples pictures will be exported.It is assumed that above-mentioned figure
The feature vector of piece 1 and picture 2 by extraction after network is respectively f1 and f2, regularization can be carried out to feature vector first
(normalization), the formula of regularization is:
Wherein, | f | indicate the mould of vector f, it is assumed that use fn1And fn2Respectively indicate f1And f2Vector after regularization, then it is two-dimentional
Euclidean distance is defined as:
Based on above-mentioned distance d, five-tuple loss can be calculated.
In one embodiment, step S220 joint classification loss and five-tuple loss to pre-trained baseline network into
Row tuning may further include:By pre-provisioning request and five samples pictures for sequentially inputting five-tuple;Based on the reference net
Network calculates Classification Loss for the predicted vector of every samples pictures output;Described five are directed to based on the baseline network
The feature vector of samples pictures output calculates five-tuple loss;And it is based on Classification Loss calculated and five-tuple calculated
Loss of the final loss of costing bio disturbance to identify network again as the pedestrian.
Joint classification loss and five-tuple loss recited above are described to pre-trained baseline network below with reference to Fig. 4
Carry out the example process of tuning.
It is as shown in Figure 4, by above-mentioned five-tuple samples pictures (including positive sample 1, positive sample 2, negative sample 1, negative sample 21
And negative sample 22) be input to it is trained after baseline network.Herein, for the reference net in the pre-training stage with step S210
Network is mutually distinguished, and the baseline network in pre-training stage is named as ID network (IDNet), and by the network in step S220 tuning stage
It is named as five-tuple-ID network (Quintuplet-IDNet), it should be understood, however, that actually the two stages are the same networks
Positive sample 1, positive sample 2, negative sample 1, negative sample 21 and negative sample 22 are input to by structure as shown in Figure 4
Quintuplet-IDNet, namely it is input to IDNet.
After above-mentioned samples pictures are input to Quintuplet-IDNet, for every samples pictures, Fc layers will output
Corresponding feature vector, Softmax layers will export corresponding predicted vector.For example, as shown in figure 4, with positive sample
1, positive sample 2, negative sample 1, negative sample 21 and the corresponding feature vector of negative sample 22 are respectively feature 1, feature
2, feature 3, feature 4 and feature 5;With positive sample 1, positive sample 2, negative sample 1, negative sample 21 and negative sample
22 corresponding predicted vectors are respectively ID1, ID2, ID3, ID4 and ID5.
Then, Classification Loss can be calculated based on the predicted vector for every samples pictures, calculation method is similar to
Described in step S210.Herein, due to input five pictures, so final Classification Loss can be this five picture
The average value of Classification Loss.Then, five-tuple loss, calculation method can be calculated based on the feature vector of five samples pictures
As described above.
Finally, can based on the final loss of Classification Loss calculated and five-tuple costing bio disturbance calculated using as
Final pedestrian identifies the loss of network again.Illustratively, the final loss is the Classification Loss calculated and institute
The weighted sum for stating five-tuple loss calculated, is expressed as:
Loss=λ lID+(1-λ)lqt
Wherein, λ is the weight parameter of 0~1 range, can voluntarily be adjusted.Illustratively, 0.5 can be set by λ.
The IDNet pre-trained come tuning is lost by the loss of above-mentioned joint classification and five-tuple, after tuning
Quintuplet-IDNet identifies that network identifies again for pedestrian as final pedestrian again.
The above-mentioned pedestrian completed based on training identifies network again, when input one opens picture probe to be checked and wait search
After the pedestrian image collection gallery of rope, it can identify that again the propagated forward of network obtains every by above-mentioned trained pedestrian
The feature vector of picture, through the feature vector of calculating probe picture at a distance from picture feature vector every in gallery,
It can obtain a sequencing of similarity.When the threshold value for being less than setting in gallery with probe picture minimum range, then it is assumed that
This picture (i.e. most like picture) and probe picture in gallery are the same pedestrians, and pedestrian's weight identification mission is complete
At.
Based on above description, pedestrian according to an embodiment of the present invention identifies the training method joint classification loss of network again
Be trained with range loss so that finally trained network possess simultaneously based on Classification Loss and based on range loss both
The advantages of method, can accelerate training process and improve precision;In addition, using five-tuple method, phase in range loss link
It is significant to shorten instruction than half or so can be shorten in traditional triple, improvement triple and four-tuple method, training time
Practice the time, and the inter- object distance that can further further, between class distance is zoomed out, to further increase precision.
The training method that pedestrian according to an embodiment of the present invention identifies network again is described above exemplarily.It is exemplary
Ground, pedestrian according to an embodiment of the present invention identify again network training method can with memory and processor equipment,
It is realized in device or system.
In addition, pedestrian according to an embodiment of the present invention identifies that the training method processing speed of network is fast again, it can be convenient ground
It is deployed in the mobile devices such as smart phone, tablet computer, personal computer.Alternatively, pedestrian according to an embodiment of the present invention
Identify that the training method of network can also be deployed in server end (or cloud) again.Alternatively, row according to an embodiment of the present invention
People identifies that the training method of network can also be deployed at server end (or cloud) and personal terminal with being distributed again.
Identify the training device of network again according to the pedestrian that another aspect provides below with reference to Fig. 5 description.Fig. 5 shows root
Identify the schematic block diagram of the training device 500 of network again according to the pedestrian of the embodiment of the present invention.
As shown in figure 5, pedestrian according to an embodiment of the present invention identifies that the training device 500 of network includes pre-training module again
510 and tuning module 520.The modules can execute the instruction for identifying network again above in conjunction with the pedestrian of Fig. 2 description respectively
Practice each step/function of method.Only pedestrian is identified again below the major function of each module of the training device 500 of network into
Row description, and omit the detail content having been described above.
Pre-training module 510 is used to carry out pre-training to baseline network using Classification Loss.Tuning module 520 is for combining
Classification Loss and five-tuple loss carry out tuning to pre-trained baseline network and identify network again to obtain pedestrian.
In one embodiment, pre-training module 510 can use Classification Loss and carry out pre-training network model, due to usual
It can make network fast convergence by tens repetitive exercises, therefore can be significantly using Classification Loss pre-training network model
Shorten the training time.
In one embodiment, pre-training module 510 is known as benchmark using the network model that Classification Loss carries out pre-training
Network, the subsequent evolutionary process of tuning module 520 are implemented on the baseline network after pre-training.Illustratively, benchmark
Network can be residual error network, for example, the residual error network of large-scale image identification challenge match (ImageNet) pre-training
(ResNet50).When baseline network be the residual error network when, pre-training module 510 by samples pictures be input to baseline network with
Before being trained, first samples pictures can be pre-processed.In other examples, pre-training module 510 can also use it
His customized convolutional network or other suitable networks are as baseline network, and correspondingly, pre-training module 510 can be by sample
This picture implements corresponding preprocessing process before being input to the baseline network.
In one embodiment, pre-training module 510 can be into one to baseline network progress pre-training using Classification Loss
Step includes:Samples pictures are input to baseline network;Predicted vector and institute by baseline network for samples pictures output
The label vector for stating samples pictures is compared to obtain Classification Loss;The baseline network is adjusted based on the Classification Loss
Parameter;And aforesaid operations are repeated, until classification accuracy and Classification Loss no longer change substantially.It is referred to tie above
Structure and pre-training process that Fig. 3 further understands said reference network are closed, for sake of simplicity, details are not described herein again.
In one embodiment, five-tuple is referred to according to five samples for having to the different pedestrians of summation sequential selection three
This picture.In one embodiment, the loss of 520 joint classification of tuning module and five-tuple loss are to pre-trained baseline network
Carrying out tuning may further include:By pre-provisioning request and five samples pictures for sequentially inputting five-tuple;Based on the benchmark
Network calculates Classification Loss for the predicted vector of every samples pictures output;Described five are directed to based on the baseline network
The feature vector for opening samples pictures output calculates five-tuple loss;And based on Classification Loss calculated and five yuan calculated
Loss of the final loss of group costing bio disturbance to identify network again as the pedestrian.It is referred to understand above in association with Fig. 4 and join
It closes Classification Loss and five-tuple loss carries out the process of tuning to pre-trained baseline network, for sake of simplicity, no longer superfluous herein
It states.
Based on above description, pedestrian according to an embodiment of the present invention identifies the training device joint classification loss of network again
Be trained with range loss so that finally trained network possess simultaneously based on Classification Loss and based on range loss both
The advantages of method, can accelerate training process and improve precision;In addition, using five-tuple method, phase in range loss link
It is significant to shorten instruction than half or so can be shorten in traditional triple, improvement triple and four-tuple method, training time
Practice the time, and the inter- object distance that can further further, between class distance is zoomed out, to further increase precision.
Fig. 6 shows the schematic block diagram that pedestrian according to an embodiment of the present invention identifies the training system 600 of network again.Row
People identifies that the training system 600 of network includes storage device 610 and processor 620 again.
Wherein, storage device 610 stores the training side for identifying network again for realizing pedestrian according to an embodiment of the present invention
The program code of corresponding steps in method.Program code of the processor 620 for being stored in Running storage device 610, to execute
Pedestrian according to an embodiment of the present invention identifies the corresponding steps of the training method of network again, and for realizing real according to the present invention
The pedestrian for applying example identifies the corresponding module in the training device of network again.
In one embodiment, pedestrian is made to identify the instruction of network again when said program code is run by processor 620
Practice system 600 and executes following steps:Pre-training is carried out to baseline network using Classification Loss;And joint classification loses and five yuan
Group loss carries out tuning to pre-trained baseline network and identifies network again to obtain pedestrian.
In one embodiment, pedestrian is made to identify the instruction of network again when said program code is run by processor 620
Practice the described of the execution of system 600 includes to baseline network progress pre-training using Classification Loss:Samples pictures are input to described
Baseline network;By the baseline network for the predicted vector of samples pictures output and the label vector of the samples pictures
It is compared to obtain Classification Loss;The parameter of the baseline network is adjusted based on the Classification Loss;And it is repeated
Step is stated, until classification accuracy and Classification Loss no longer change substantially.
In one embodiment, the baseline network is residual error network.
In one embodiment, before the samples pictures are input to the baseline network, to the samples pictures
Implement pretreatment operation.
In one embodiment, pedestrian is made to identify the instruction of network again when said program code is run by processor 620
The joint classification loss and five-tuple loss for practicing the execution of system 600 carry out tuning to pre-trained baseline network and include:
By pre-provisioning request and five samples pictures for sequentially inputting five-tuple;Every samples pictures are directed to based on the baseline network
The predicted vector of output calculates Classification Loss;The feature vector of five samples pictures output is directed to based on the baseline network
Calculate five-tuple loss;And based on the final loss of Classification Loss calculated and five-tuple costing bio disturbance calculated to make
Identify the loss of network again for the pedestrian.
In one embodiment, the Classification Loss calculated is being averaged for the Classification Loss of five samples pictures
Value.
In one embodiment, the five-tuple loss is defined as:
lqt=d (positive sample 1, positive sample 2)-d (negative sample 1, negative sample 21)+d (negative sample 21, negative sample 22)-d is (negative
Sample 1, positive sample 2)+a
Wherein, lqtFor five-tuple loss;Positive sample 1, positive sample 2, negative sample 1, negative sample 21 and negative sample 22 are institute
Five samples pictures are stated, and positive sample 1 and positive sample 2 are two different pictures of the first pedestrian, negative sample 1 is the second pedestrian's
Picture, negative sample 21 and negative sample 22 are two different pictures of third pedestrian;D be two pictures feature vector between away from
From;A is the constant parameter being arranged according to demand.
In one embodiment, the final loss is the Classification Loss calculated and five yuan calculated described
The weighted sum of group loss.
In addition, according to embodiments of the present invention, additionally providing a kind of storage medium, storing program on said storage
Instruction identifies network for executing the pedestrian of the embodiment of the present invention when described program instruction is run by computer or processor again
Training method corresponding steps, and identify again for realizing pedestrian according to an embodiment of the present invention in the training device of network
Corresponding module.The storage medium for example may include the storage card of smart phone, the storage unit of tablet computer, personal meter
The hard disk of calculation machine, read-only memory (ROM), Erasable Programmable Read Only Memory EPROM (EPROM), the read-only storage of portable compact disc
Any combination of device (CD-ROM), USB storage or above-mentioned storage medium.The computer readable storage medium can be
Any combination of one or more computer readable storage mediums, such as a computer readable storage medium include to utilize classification
The computer-readable program code that pre-training is carried out to baseline network is lost, another computer readable storage medium includes connection
It closes Classification Loss and five-tuple loss and tuning is carried out to obtain the calculating that pedestrian identifies network again to pre-trained baseline network
The readable program code of machine.
In one embodiment, the computer program instructions may be implemented real according to the present invention when being run by computer
The pedestrian for applying example identifies each functional module of the training device of network again, and/or can execute and implement according to the present invention
The pedestrian of example identifies the training method of network again.
In one embodiment, the computer program instructions make computer or place when being run by computer or processor
It manages device and executes following steps:Pre-training is carried out to baseline network using Classification Loss;And joint classification loss and five-tuple damage
It loses and network is identified again to obtain pedestrian to pre-trained baseline network progress tuning.
In one embodiment, the computer program instructions make computer or place when being run by computer or processor
Manage the described of device execution includes to baseline network progress pre-training using Classification Loss:Samples pictures are input to the reference net
Network;Label vector by the baseline network for the predicted vector and the samples pictures of samples pictures output compares
Compared with to obtain Classification Loss;The parameter of the baseline network is adjusted based on the Classification Loss;And above-mentioned steps are repeated,
Until classification accuracy and Classification Loss no longer change substantially.
In one embodiment, the baseline network is residual error network.
In one embodiment, before the samples pictures are input to the baseline network, to the samples pictures
Implement pretreatment operation.
In one embodiment, the computer program instructions make computer or place when being run by computer or processor
The joint classification loss and five-tuple loss that reason device executes carry out tuning to pre-trained baseline network and include:By predetermined
It is required that and sequentially inputting five samples pictures of five-tuple;It is exported based on the baseline network for every samples pictures
Predicted vector calculates Classification Loss;Feature vector based on the baseline network for five samples pictures output calculates five
Tuple loss;And based on the final loss of Classification Loss calculated and five-tuple costing bio disturbance calculated using as described
Pedestrian identifies the loss of network again.
In one embodiment, the Classification Loss calculated is being averaged for the Classification Loss of five samples pictures
Value.
In one embodiment, the five-tuple loss is defined as:
lqt=d (positive sample 1, positive sample 2)-d (negative sample 1, negative sample 21)+d (negative sample 21, negative sample 22)-d is (negative
Sample 1, positive sample 2)+a
Wherein, lqtFor five-tuple loss;Positive sample 1, positive sample 2, negative sample 1, negative sample 21 and negative sample 22 are institute
Five samples pictures are stated, and positive sample 1 and positive sample 2 are two different pictures of the first pedestrian, negative sample 1 is the second pedestrian's
Picture, negative sample 21 and negative sample 22 are two different pictures of third pedestrian;D be two pictures feature vector between away from
From;A is the constant parameter being arranged according to demand.
In one embodiment, the final loss is the Classification Loss calculated and five yuan calculated described
The weighted sum of group loss.
Pedestrian according to an embodiment of the present invention identifies that each module in the training device of network can be by according to this hair again
The pedestrian of bright embodiment identifies that the processor of the electronic equipment of the training of network runs the computer journey stored in memory again
Sequence instructs to realize, or can be in the computer readable storage medium of computer program product according to an embodiment of the present invention
The realization when computer instruction of storage is run by computer.
Pedestrian according to an embodiment of the present invention identifies training method, device, system and the storage medium joint of network again
Classification Loss and range loss are trained, so that finally trained network possesses simultaneously based on Classification Loss and based on distance damage
The advantages of losing both methods, can accelerate training process and improve precision;In addition, using five-tuple in range loss link
Method can shorten to half or so compared to traditional triple, improvement triple and four-tuple method, training time, show
It writes and shortens the training time, and the inter- object distance that can further further, between class distance is zoomed out, to further increase precision.
Pedestrian according to an embodiment of the present invention is described to example above to identify the training method of network again, device, be
System and storage medium.The present invention also provides a kind of pedestrians again recognition methods, and pedestrian described above is used to identify network again
Training method training made of pedestrian identify again network carry out pedestrian identify again.The present invention also provides a kind of pedestrians to identify again
Device is used to implement pedestrian recognition methods again.The present invention also provides a kind of pedestrian weight identifying systems comprising storage dress
It sets and processor, is stored with the computer program run by the processor on the storage device, the computer program exists
Pedestrian recognition methods again is executed when being run by the processor.The present invention also provides a kind of storage medium, the storage is situated between
Computer program is stored in matter, the computer program executes pedestrian recognition methods again at runtime.Those skilled in the art
Member can identify that training method, device, system and the storage of network are situated between again based on pedestrian according to an embodiment of the present invention above-mentioned
Matter understands pedestrian according to an embodiment of the present invention recognition methods, device, system and storage medium again, for sake of simplicity, herein no longer
It repeats.
Although describing example embodiment by reference to attached drawing here, it should be understood that above example embodiment are only exemplary
, and be not intended to limit the scope of the invention to this.Those of ordinary skill in the art can carry out various changes wherein
And modification, it is made without departing from the scope of the present invention and spiritual.All such changes and modifications are intended to be included in appended claims
Within required the scope of the present invention.
Those of ordinary skill in the art may be aware that list described in conjunction with the examples disclosed in the embodiments of the present disclosure
Member and algorithm steps can be realized with the combination of electronic hardware or computer software and electronic hardware.These functions are actually
It is implemented in hardware or software, the specific application and design constraint depending on technical solution.Professional technician
Each specific application can be used different methods to achieve the described function, but this realization is it is not considered that exceed
The scope of the present invention.
In several embodiments provided herein, it should be understood that disclosed device and method can pass through it
Its mode is realized.For example, apparatus embodiments described above are merely indicative, for example, the division of the unit, only
Only a kind of logical function partition, there may be another division manner in actual implementation, such as multiple units or components can be tied
Another equipment is closed or is desirably integrated into, or some features can be ignored or not executed.
In the instructions provided here, numerous specific details are set forth.It is to be appreciated, however, that implementation of the invention
Example can be practiced without these specific details.In some instances, well known method, structure is not been shown in detail
And technology, so as not to obscure the understanding of this specification.
Similarly, it should be understood that in order to simplify the present invention and help to understand one or more of the various inventive aspects,
To in the description of exemplary embodiment of the present invention, each feature of the invention be grouped together into sometimes single embodiment, figure,
Or in descriptions thereof.However, the method for the invention should not be construed to reflect following intention:It is i.e. claimed
The present invention claims features more more than feature expressly recited in each claim.More precisely, such as corresponding power
As sharp claim reflects, inventive point is that the spy of all features less than some disclosed single embodiment can be used
Sign is to solve corresponding technical problem.Therefore, it then follows thus claims of specific embodiment are expressly incorporated in this specific
Embodiment, wherein each, the claims themselves are regarded as separate embodiments of the invention.
It will be understood to those skilled in the art that any combination pair can be used other than mutually exclusive between feature
All features disclosed in this specification (including adjoint claim, abstract and attached drawing) and so disclosed any method
Or all process or units of equipment are combined.Unless expressly stated otherwise, this specification (is wanted including adjoint right
Ask, make a summary and attached drawing) disclosed in each feature can be replaced with an alternative feature that provides the same, equivalent, or similar purpose.
In addition, it will be appreciated by those of skill in the art that although some embodiments described herein include other embodiments
In included certain features rather than other feature, but the combination of the feature of different embodiments mean it is of the invention
Within the scope of and form different embodiments.For example, in detail in the claims, embodiment claimed it is one of any
Can in any combination mode come using.
Various component embodiments of the invention can be implemented in hardware, or to run on one or more processors
Software module realize, or be implemented in a combination thereof.It will be understood by those of skill in the art that can be used in practice
Microprocessor or digital signal processor (DSP) realize some or all of some modules according to an embodiment of the present invention
Function.The present invention is also implemented as some or all program of device (examples for executing method as described herein
Such as, computer program and computer program product).It is such to realize that program of the invention can store in computer-readable medium
On, or may be in the form of one or more signals.Such signal can be downloaded from an internet website to obtain, or
Person is provided on the carrier signal, or is provided in any other form.
It should be noted that the above-mentioned embodiments illustrate rather than limit the invention, and ability
Field technique personnel can be designed alternative embodiment without departing from the scope of the appended claims.In the claims,
Any reference symbol between parentheses should not be configured to limitations on claims.Word "comprising" does not exclude the presence of not
Element or step listed in the claims.Word "a" or "an" located in front of the element does not exclude the presence of multiple such
Element.The present invention can be by means of including the hardware of several different elements and being come by means of properly programmed computer real
It is existing.In the unit claims listing several devices, several in these devices can be through the same hardware branch
To embody.The use of word first, second, and third does not indicate any sequence.These words can be explained and be run after fame
Claim.
The above description is merely a specific embodiment or to the explanation of specific embodiment, protection of the invention
Range is not limited thereto, and anyone skilled in the art in the technical scope disclosed by the present invention, can be easily
Expect change or replacement, should be covered by the protection scope of the present invention.Protection scope of the present invention should be with claim
Subject to protection scope.
Claims (20)
1. the training method that a kind of pedestrian identifies network again, which is characterized in that the training method includes:
Pre-training is carried out to baseline network using Classification Loss;And
Joint classification loss and five-tuple loss carry out tuning to pre-trained baseline network and identify network again to obtain pedestrian.
2. training method according to claim 1, which is characterized in that described to be carried out in advance using Classification Loss to baseline network
Training includes:
Samples pictures are input to the baseline network;
The baseline network is carried out for the predicted vector of samples pictures output and the label vector of the samples pictures
Compare to obtain Classification Loss;
The parameter of the baseline network is adjusted based on the Classification Loss;And
Above-mentioned steps are repeated, until classification accuracy and Classification Loss no longer change substantially.
3. training method according to claim 2, which is characterized in that the baseline network is residual error network.
4. training method according to claim 3, which is characterized in that the samples pictures are being input to the reference net
Before network, pretreatment operation is implemented to the samples pictures.
5. training method according to claim 1, which is characterized in that the joint classification loss and five-tuple loss are to warp
The baseline network of pre-training carries out tuning:
By pre-provisioning request and five samples pictures for sequentially inputting five-tuple;
Predicted vector based on the baseline network for every samples pictures output calculates Classification Loss;
Five-tuple loss is calculated for the feature vector of five samples pictures output based on the baseline network;And
Based on the final loss of Classification Loss calculated and five-tuple costing bio disturbance calculated to know as the pedestrian again
The loss of other network.
6. training method according to claim 5, which is characterized in that the Classification Loss calculated is five samples
The average value of the Classification Loss of this picture.
7. training method according to claim 5, which is characterized in that the five-tuple loss is defined as:
lqt=d (positive sample 1, positive sample 2)-d (negative sample 1, negative sample 21)+d (negative sample 21, negative sample 22)-d (negative sample
1, positive sample 2)+a
Wherein, lqtFor five-tuple loss;Positive sample 1, positive sample 2, negative sample 1, negative sample 21 and negative sample 22 are described five
Samples pictures are opened, and positive sample 1 and positive sample 2 are two different pictures of the first pedestrian, negative sample 1 is the figure of the second pedestrian
Piece, negative sample 21 and negative sample 22 are two different pictures of third pedestrian;D be two pictures feature vector between away from
From;A is the constant parameter being arranged according to demand.
8. the training method according to any one of claim 5-7, which is characterized in that the final loss is described
The weighted sum of Classification Loss calculated and the five-tuple loss calculated.
9. the training device that a kind of pedestrian identifies network again, which is characterized in that the training device includes:
Pre-training module, for carrying out pre-training to baseline network using Classification Loss;And
Tuning module carries out tuning to pre-trained baseline network for joint classification loss and five-tuple loss to be gone
People identifies network again.
10. training device according to claim 9, which is characterized in that the pre-training module is to the baseline network
Pre-training further comprises:
Samples pictures are input to the baseline network;
The baseline network is carried out for the predicted vector of samples pictures output and the label vector of the samples pictures
Compare to obtain Classification Loss;
The parameter of the baseline network is adjusted based on the Classification Loss;And
Aforesaid operations are repeated, until classification accuracy and Classification Loss no longer change substantially.
11. training device according to claim 10, which is characterized in that the baseline network is residual error network.
12. training device according to claim 11, which is characterized in that the pre-training module is also used to:Will be described
Samples pictures are input to before the baseline network, implement pretreatment operation to the samples pictures.
13. training device according to claim 9, which is characterized in that the tuning module is to pre-trained reference net
The tuning of network includes:
By pre-provisioning request and five samples pictures for sequentially inputting five-tuple;
Predicted vector based on the baseline network for every samples pictures output calculates Classification Loss;
Five-tuple loss is calculated for the feature vector of five samples pictures output based on the baseline network;And
Based on the final loss of Classification Loss calculated and five-tuple costing bio disturbance calculated to know as the pedestrian again
The loss of other network.
14. training device according to claim 13, which is characterized in that the Classification Loss calculated is described five
The average value of the Classification Loss of samples pictures.
15. training device according to claim 13, which is characterized in that the five-tuple loss is defined as:
lqt=d (positive sample 1, positive sample 2)-d (negative sample 1, negative sample 21)+d (negative sample 21, negative sample 22)-d (negative sample
1, positive sample 2)+a
Wherein, lqtFor five-tuple loss;Positive sample 1, positive sample 2, negative sample 1, negative sample 21 and negative sample 22 are described five
Samples pictures are opened, and positive sample 1 and positive sample 2 are two different pictures of the first pedestrian, negative sample 1 is the figure of the second pedestrian
Piece, negative sample 21 and negative sample 22 are two different pictures of third pedestrian;D be two pictures feature vector between away from
From;A is the constant parameter being arranged according to demand.
16. training device described in any one of 3-15 according to claim 1, which is characterized in that the final loss is
The weighted sum of the Classification Loss calculated and the five-tuple loss calculated.
17. a kind of pedestrian recognition methods again, which is characterized in that the pedestrian again recognition methods using according to claim 1 in -8
Described in any item pedestrians identify again network training method training made of pedestrian identify again network carry out pedestrian identify again.
18. a kind of pedestrian weight identification device, which is characterized in that pedestrian's weight identification device is for implementing according to claim 17
Pedestrian recognition methods again.
19. a kind of computing system, which is characterized in that the system comprises storage device and processor, deposited on the storage device
The computer program run by the processor is contained, the computer program executes such as right when being run by the processor
It is required that pedestrian described in any one of 1-8 identifies the training method of network or executes pedestrian's weight as claimed in claim 17 again
Recognition methods.
20. a kind of storage medium, which is characterized in that be stored with computer program, the computer program on the storage medium
The pedestrian as described in any one of claim 1-8 is executed at runtime to identify the training method of network again or execute such as right
It is required that the recognition methods again of pedestrian described in 17.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710906719.5A CN108875487B (en) | 2017-09-29 | 2017-09-29 | Training of pedestrian re-recognition network and pedestrian re-recognition based on training |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710906719.5A CN108875487B (en) | 2017-09-29 | 2017-09-29 | Training of pedestrian re-recognition network and pedestrian re-recognition based on training |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108875487A true CN108875487A (en) | 2018-11-23 |
CN108875487B CN108875487B (en) | 2021-06-15 |
Family
ID=64325776
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710906719.5A Active CN108875487B (en) | 2017-09-29 | 2017-09-29 | Training of pedestrian re-recognition network and pedestrian re-recognition based on training |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108875487B (en) |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110321965A (en) * | 2019-07-10 | 2019-10-11 | 腾讯科技(深圳)有限公司 | The method and device that the training method of object weight identification model, object identify again |
CN110688888A (en) * | 2019-08-02 | 2020-01-14 | 浙江省北大信息技术高等研究院 | Pedestrian attribute identification method and system based on deep learning |
CN110728216A (en) * | 2019-09-27 | 2020-01-24 | 西北工业大学 | Unsupervised pedestrian re-identification method based on pedestrian attribute adaptive learning |
CN111597876A (en) * | 2020-04-01 | 2020-08-28 | 浙江工业大学 | Cross-modal pedestrian re-identification method based on difficult quintuple |
CN111611846A (en) * | 2020-03-31 | 2020-09-01 | 北京迈格威科技有限公司 | Pedestrian re-identification method and device, electronic equipment and storage medium |
CN111738172A (en) * | 2020-06-24 | 2020-10-02 | 中国科学院自动化研究所 | Cross-domain target re-identification method based on feature counterstudy and self-similarity clustering |
CN111814655A (en) * | 2020-07-03 | 2020-10-23 | 浙江大华技术股份有限公司 | Target re-identification method, network training method thereof and related device |
CN114067356A (en) * | 2021-10-21 | 2022-02-18 | 电子科技大学 | Pedestrian re-identification method based on joint local guidance and attribute clustering |
US11948391B2 (en) | 2020-11-25 | 2024-04-02 | Boe Technology Group Co., Ltd. | Model training method and apparatus, electronic device and readable storage medium |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105975959A (en) * | 2016-06-14 | 2016-09-28 | 广州视源电子科技股份有限公司 | Face feature extraction modeling and face recognition method and device based on neural network |
CN106203533A (en) * | 2016-07-26 | 2016-12-07 | 厦门大学 | The degree of depth based on combined training study face verification method |
CN106250870A (en) * | 2016-08-16 | 2016-12-21 | 电子科技大学 | A kind of pedestrian's recognition methods again combining local and overall situation similarity measurement study |
CN106778527A (en) * | 2016-11-28 | 2017-05-31 | 中通服公众信息产业股份有限公司 | A kind of improved neutral net pedestrian recognition methods again based on triple losses |
WO2017123477A1 (en) * | 2016-01-11 | 2017-07-20 | Flir Systems, Inc. | Vehicle based radar upsampling |
CN107038448A (en) * | 2017-03-01 | 2017-08-11 | 中国科学院自动化研究所 | Target detection model building method |
CN107103281A (en) * | 2017-03-10 | 2017-08-29 | 中山大学 | Face identification method based on aggregation Damage degree metric learning |
-
2017
- 2017-09-29 CN CN201710906719.5A patent/CN108875487B/en active Active
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2017123477A1 (en) * | 2016-01-11 | 2017-07-20 | Flir Systems, Inc. | Vehicle based radar upsampling |
CN105975959A (en) * | 2016-06-14 | 2016-09-28 | 广州视源电子科技股份有限公司 | Face feature extraction modeling and face recognition method and device based on neural network |
CN106203533A (en) * | 2016-07-26 | 2016-12-07 | 厦门大学 | The degree of depth based on combined training study face verification method |
CN106250870A (en) * | 2016-08-16 | 2016-12-21 | 电子科技大学 | A kind of pedestrian's recognition methods again combining local and overall situation similarity measurement study |
CN106778527A (en) * | 2016-11-28 | 2017-05-31 | 中通服公众信息产业股份有限公司 | A kind of improved neutral net pedestrian recognition methods again based on triple losses |
CN107038448A (en) * | 2017-03-01 | 2017-08-11 | 中国科学院自动化研究所 | Target detection model building method |
CN107103281A (en) * | 2017-03-10 | 2017-08-29 | 中山大学 | Face identification method based on aggregation Damage degree metric learning |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110321965A (en) * | 2019-07-10 | 2019-10-11 | 腾讯科技(深圳)有限公司 | The method and device that the training method of object weight identification model, object identify again |
CN110321965B (en) * | 2019-07-10 | 2021-06-18 | 腾讯科技(深圳)有限公司 | Training method of object re-recognition model, and object re-recognition method and device |
CN110688888B (en) * | 2019-08-02 | 2022-08-05 | 杭州未名信科科技有限公司 | Pedestrian attribute identification method and system based on deep learning |
CN110688888A (en) * | 2019-08-02 | 2020-01-14 | 浙江省北大信息技术高等研究院 | Pedestrian attribute identification method and system based on deep learning |
CN110728216A (en) * | 2019-09-27 | 2020-01-24 | 西北工业大学 | Unsupervised pedestrian re-identification method based on pedestrian attribute adaptive learning |
CN111611846A (en) * | 2020-03-31 | 2020-09-01 | 北京迈格威科技有限公司 | Pedestrian re-identification method and device, electronic equipment and storage medium |
CN111597876A (en) * | 2020-04-01 | 2020-08-28 | 浙江工业大学 | Cross-modal pedestrian re-identification method based on difficult quintuple |
CN111738172A (en) * | 2020-06-24 | 2020-10-02 | 中国科学院自动化研究所 | Cross-domain target re-identification method based on feature counterstudy and self-similarity clustering |
CN111738172B (en) * | 2020-06-24 | 2021-02-12 | 中国科学院自动化研究所 | Cross-domain target re-identification method based on feature counterstudy and self-similarity clustering |
CN111814655A (en) * | 2020-07-03 | 2020-10-23 | 浙江大华技术股份有限公司 | Target re-identification method, network training method thereof and related device |
CN111814655B (en) * | 2020-07-03 | 2023-09-01 | 浙江大华技术股份有限公司 | Target re-identification method, network training method thereof and related device |
US11948391B2 (en) | 2020-11-25 | 2024-04-02 | Boe Technology Group Co., Ltd. | Model training method and apparatus, electronic device and readable storage medium |
CN114067356A (en) * | 2021-10-21 | 2022-02-18 | 电子科技大学 | Pedestrian re-identification method based on joint local guidance and attribute clustering |
Also Published As
Publication number | Publication date |
---|---|
CN108875487B (en) | 2021-06-15 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108875487A (en) | Pedestrian is identified the training of network again and is identified again based on its pedestrian | |
CN108197532B (en) | The method, apparatus and computer installation of recognition of face | |
CN108875486A (en) | Recongnition of objects method, apparatus, system and computer-readable medium | |
CN108875932A (en) | Image-recognizing method, device and system and storage medium | |
WO2018227800A1 (en) | Neural network training method and device | |
CN110309706A (en) | Face critical point detection method, apparatus, computer equipment and storage medium | |
CN106203305A (en) | Human face in-vivo detection method and device | |
CN108337000A (en) | Automated process for being transformed into lower accuracy data format | |
CN109086811A (en) | Multi-tag image classification method, device and electronic equipment | |
CN110532884A (en) | Pedestrian recognition methods, device and computer readable storage medium again | |
CN108229419A (en) | For clustering the method and apparatus of image | |
CN109934173A (en) | Expression recognition method, device and electronic equipment | |
CN108875492A (en) | Face datection and crucial independent positioning method, device, system and storage medium | |
CN106096028A (en) | Historical relic indexing means based on image recognition and device | |
CN110009614A (en) | Method and apparatus for output information | |
CN113128478B (en) | Model training method, pedestrian analysis method, device, equipment and storage medium | |
CN110210513A (en) | Data classification method, device and terminal device | |
CN110163205A (en) | Image processing method, device, medium and calculating equipment | |
CN110298240A (en) | A kind of user vehicle recognition methods, device, system and storage medium | |
CN110347724A (en) | Abnormal behaviour recognition methods, device, electronic equipment and medium | |
CN108875517A (en) | Method for processing video frequency, device and system and storage medium | |
CN107958247A (en) | Method and apparatus for facial image identification | |
CN110516734A (en) | A kind of image matching method, device, equipment and storage medium | |
CN106778910A (en) | Deep learning system and method based on local training | |
CN110225368A (en) | A kind of video locating method, device and electronic equipment |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |