CN113657254B - Pedestrian re-identification domain adaptation method based on reliable value sample and new identity sample mining - Google Patents
Pedestrian re-identification domain adaptation method based on reliable value sample and new identity sample mining Download PDFInfo
- Publication number
- CN113657254B CN113657254B CN202110935835.6A CN202110935835A CN113657254B CN 113657254 B CN113657254 B CN 113657254B CN 202110935835 A CN202110935835 A CN 202110935835A CN 113657254 B CN113657254 B CN 113657254B
- Authority
- CN
- China
- Prior art keywords
- sample
- target domain
- clustering
- value
- samples
- 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.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
Abstract
The invention discloses a pedestrian re-recognition domain adaptation method based on reliable value sample and new identity sample mining. Sample selection is performed by measuring the reliability and information quantity of the sample pseudo tag on one hand, and potential new identity samples are created for domain adaptation by a sample mixing mode on the other hand. The invention designs and uses the proper training sample for the adaptation and optimization of the model to the target domain distribution, thereby improving the robustness and generalization of the target domain model.
Description
Technical Field
The invention belongs to the technical field of intelligent recognition, and particularly relates to a pedestrian re-recognition domain adaptation method based on reliable value sample and new identity sample mining. Deep learning techniques are involved in the algorithm design and model training section.
Background
The task of pedestrian re-recognition is to search for pedestrian targets in camera B where there is no overlapping area of camera a, and to re-find pedestrian targets that appear in camera a. As a current important research direction and research hotspot, pedestrian re-recognition has wide application in the fields of intelligent monitoring, smart cities, public security control, criminal investigation and the like, such as pedestrian tracking and behavior analysis across cameras, picture retrieval and inquiry of suspected or interested people and the like.
With the rapid development of deep learning technology and the strong learning fitting capability of convolutional neural networks, the pedestrian re-recognition algorithm based on the identity tag has obtained very high recognition accuracy and first hit rate on the mainstream data set, and even surpasses the recognition capability of human eyes. However, labeling of pedestrian identity categories is a very tedious and labor-intensive project, and in a real monitoring scene, massive new pedestrian data are generated at any time, so that artificial labeling of such massive data becomes increasingly infeasible, and therefore, how to learn and adapt unlabeled data distributed in an unknown scene based on currently existing labeled data and trained models is gradually focused by more researchers.
The existing high-performance pedestrian re-recognition domain adaptation model mostly utilizes a source domain pre-trained model to cluster and distribute pseudo tags to target domain data, updates the model according to the pseudo tags of the target domain data, and repeats the process until convergence. However, due to the instability of early models and the inadaptability to the target domain, the pseudo tags assigned by the models may contain a lot of noise, and the erroneous pseudo tags may mislead the learning of the models, resulting in performance degradation. In addition, because different identity categories are included in training and testing in the task of re-identifying pedestrians, the model needs to process the newly-appearing identity categories of the pedestrians in the testing set besides domain adaptation, and higher requirements are set for generalization of the model.
Disclosure of Invention
The invention aims to provide a pedestrian re-identification domain adaptation method based on reliable value samples and new identity samples for mining, aiming at the defects of the prior art.
The aim of the invention is realized by the following technical scheme: a pedestrian re-identification domain adaptation method based on reliable value sample and new identity sample mining comprises the following steps:
step one: a pedestrian target re-recognition model F is trained on the source domain data.
Step two: clustering by using the model F as target domain data, and distributing pseudo tags according to a clustering result;
step three: selecting samples with high clustering reliability and high clustering value from the target domain data according to the clustering result, and forming a reliable and valuable sample set { X (X) in the target domain by corresponding pseudo labels r ,Y r -a }; concrete embodimentsObtained by the following substeps:
(3.1) for each sample in the target domain, calculating the clustering reliability according to the clustering result, and selecting samples with high reliability and corresponding pseudo tags to form a reliable sample set { X } c ,Y c -a }; the clustering reliability is the ratio of the similarity between the feature obtained by each target domain sample according to the model F and the closest clustering center to the feature and the second closest clustering center, and the higher the ratio is, the higher the reliability is.
(3.2) for each sample in the target domain, calculating the clustering value according to the clustering result, and selecting samples with high value and corresponding pseudo tags to form a value sample set { X } v ,Y v -a }; the clustering value is dissimilarity between the characteristics obtained by each target domain sample according to the model F and the characteristics of other target domain samples, and the higher the dissimilarity is, the higher the value is. The other target domain sample features are kth other target domain sample features which are sequenced from high to low according to the feature similarity of the target domain sample and other target domain samples, and the dissimilarity is the reciprocal of the similarity.
(3.3) in the reliable sample set { X ] c ,Y c Sum value sample set { X } v ,Y v Finding co-occurring samples in the target domain, obtaining a reliable and valuable set of samples { X } r ,Y r }。
Step four: using source domain data { X ] s ,Y s Reliable and valuable sample set { X } in the target domain r ,Y r Generating a new identity sample set { X } n ,Y n }；
Step five: utilizing reliable and valuable sample sets { X in the target domain r ,Y r Sum of new identity sample set { X } n ,Y n Training an update model F;
step six: and repeating the second step to the fifth step until the model converges to obtain a final target domain model F.
Further, the cluster reliability and reliable sample set { X in (3.1) c ,Y c -calculated by:
(3.1.1) calculating a cluster center c of each cluster, which is obtained by averaging the characteristics of all samples in the clusters;
(3.1.2) for each target field sample { x } t ,y t Computing features F by model F t Find the nearest cluster center c i And the second closest cluster center c o ；
(3.1.3) calculating the target Domain sample { x } using t ,y t Cluster reliability of }:
wherein the method comprises the steps ofRepresenting the similarity between vector a and vector b;
(3.1.4) calculating the clustering reliability of all samples in the target domain, and averaging to obtain the mean clustering reliability tau g ；
(3.1.5) selecting samples with the clustering reliability greater than the mean clustering reliability, namely meeting R g ＞τ g Form a reliable sample set { X } c ,Y c }。
Further, clustering value and value sample set { X ] in (3.2) v ,Y v -calculated by:
(3.2.1) for each target field sample { x } t ,y t Computing features F by model F t ；
(3.2.2) calculating the feature similarity of each target domain sample and other target domain samples and selecting the sample feature f of the k th bit of the ordered list according to the order from high to low k And obtain its dissimilarityCluster value R as the sample l The method comprises the steps of carrying out a first treatment on the surface of the k is a positive integer.
(3.2.3) computing all of the target DomainClustering value of samples and averaging to obtain average clustering value tau l ；
(3.2.4) selecting samples with cluster value greater than the average value cluster value, namely meeting R l ＞τ l Form a value sample set { X } v ,Y v }。
Further, in the fourth step, a new identity sample set { X } n ,Y n -obtained by:
(4.1) randomly selecting one sample { x } in the source domain data s ,y s }；
(4.2) reliable and valuable sample set { X over target field r ,Y r Sequentially selecting one sample { x } r ,y r }；
(4.3) mixing the randomly selected samples and their labels in (4.1) and (4.2) respectively to obtain a new identity sample { x }, using the following formula n ,y n }：
x n ＝αx s +(1-α)x r
y n ＝Softmax(αy s +(1-α)y r )
Where α is a randomly sampled variable.
(4.4) repeating steps (4.1) to (4.3) for a reliable and valuable sample set { X over a target domain r ,Y r Each sample in the sequence generates a new identity sample to form a new identity sample set { X }, and n ,Y n }。
further, in (4.3), α is a random sample from the distribution Beta (0.5 ).
The beneficial effects of the invention are as follows: the invention digs and samples reliable and valuable samples in the target domain through a reasonable design strategy, is used for domain adaptation training of the model, and effectively avoids negative effects caused by pseudo tag noise. Meanwhile, in order to improve the expression of the model on the target domain test data, the invention creates a possible new identity sample in a sample mixing mode, and effectively enhances the identity discrimination and generalization of the model for unknown identity pedestrians.
Drawings
FIG. 1 is a flow chart of steps of a pedestrian re-identification domain adaptation method based on reliable sample and new identity sample mining.
Detailed Description
The present invention will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.
On the contrary, the invention is intended to cover any alternatives, modifications, equivalents, and variations as may be included within the spirit and scope of the invention as defined by the appended claims. Further, in the following detailed description of the present invention, certain specific details are set forth in order to provide a better understanding of the present invention. The present invention will be fully understood by those skilled in the art without the details described herein.
The invention provides a pedestrian re-identification domain adaptation method based on reliable value sample and new identity sample excavation, which comprises the following steps:
step one: a pedestrian target re-recognition model F is trained on the source domain data. Specifically, in this embodiment, a ResNet-50 network is used as the structure of F.
Step two: and clustering the target domain data by using the model F, and distributing pseudo tags according to a clustering result. Specifically, the target domain data is input into a model F to obtain the characteristics of each sample, then clustering is carried out according to the characteristics, and pseudo tags are distributed.
Step three: selecting samples with high clustering reliability and high clustering value from the target domain data according to the clustering result, and forming a reliable and valuable sample set { X (X) in the target domain by corresponding pseudo labels r ,Y r }；
Reliable and valuable sample set { X over target domain in step three above r ,Y r Acquisition, specifically:
(3.1) for each sample in the target domainThe method comprises the steps of calculating the clustering reliability according to a clustering result, selecting samples with high reliability and corresponding pseudo tags to form a reliable sample set { X } c ,Y c -a }; specifically:
(3.1.1) calculating a cluster center c of each cluster, which is obtained by averaging the characteristics of all samples in the clusters;
(3.1.2) for each target field sample { x } t ,y t Computing features F by model F t Find the nearest cluster center c i And the second closest cluster center c o ；
(3.1.3) calculating the target Domain sample { x } using t ,y t Cluster reliability of }:
wherein the method comprises the steps ofRepresenting the similarity between vector b and vector a;
(3.1.4) calculating the clustering reliability of all samples in the target domain, and averaging to obtain the mean clustering reliability tau g ；
(3.1.5) selecting samples with the clustering reliability greater than the mean clustering reliability, namely meeting R g ＞τ g Form a reliable sample set { X } c ,Y c }。
(3.2) for each sample in the target domain, calculating the clustering value according to the clustering result, and selecting samples with high value and corresponding pseudo tags to form a value sample set { X } v ,Y v -a }; specifically:
(3.2.1) for each target field sample { x } t ,y t Computing features F by model F t ；
(3.2.2) calculating the feature similarity of each target domain sample and other target domain samples and selecting the sample feature f of the k th bit of the ordered list according to the order from high to low k And obtain its dissimilarityCluster value R as the sample l The method comprises the steps of carrying out a first treatment on the surface of the k is a positive integer, and is set to k=10 according to the final experimental effect selection.
(3.2.3) calculating the cluster value of all samples in the target domain and averaging to obtain the mean cluster value τ l ；
(3.2.4) selecting samples with cluster value greater than the average value cluster value, namely meeting R l ＞τ l Form a value sample set { X } v ,Y v }。
(3.3) in the reliable sample set { X ] c ,Y c Sum value sample set { X } v ,Y v Finding co-occurring samples in the target domain, obtaining a reliable and valuable set of samples { X } r ,Y r }。
Step four: reliable and valuable sample set { X over source domain data and target domain r ,Y r Generation of a new identity sample set X n ,Y n -a }; specifically:
(4.1) randomly selecting one sample { x } in the source domain data s ,y s }；
(4.2) reliable and valuable sample set { X over target field r ,Y r Sequentially selecting one sample { x } r ,y r }；
(4.3) mixing the randomly selected samples and their labels in (4.1) and (4.2) respectively to obtain a new identity sample { x }, using the following formula n ,y n }：
x n ＝αx s +(1-α)x r
y n ＝Softmax(αy s +(1-α)y r )
Where α is a variable randomly sampled from the distribution Beta (0.5 ). Softmax is a Softmax function.
(4.4) repeating steps (4.1) to (4.3) for a reliable and valuable sample set { X over a target domain r ,Y r Each sample in the sequence generates a new identity sample,constitute a new identity sample set { X } n ,Y n }。
Step five: utilizing reliable and valuable sample sets { X in the target domain r ,Y r Sum of new identity sample set { X } n ,Y n Training update model F.
Specifically, a cross entropy classification loss function is adopted, and training and updating are carried out on the model F according to the pseudo tag corresponding to the target domain sample:
wherein N represents the number of sample pictures in a training batch, Y i And (3) representing the identity class corresponding to the ith sample, and p (i) representing the prediction probability of the identity class.
Step six: and repeating the second step to the fifth step until the model converges to obtain a final target domain model F for the pedestrian target re-recognition task test on the target domain.
Table 1 is a comparison of performance improvement of the reliable value sample selection strategy and the new identity sample mining strategy proposed by the present invention for the task of pedestrian re-identification domain adaptation. The experiment uses Resnet-50 as a pedestrian target re-recognition model F to test on two data sets of a Market-1501 and a DukeMTMC-reiD respectively, and the DukeMTMC-reiD is used as a source domain data set (DukeMTMC-reiD- > Market-1501) when the Market-1501 is tested; in the performance test of DukeMTMC-reiD database, market-1501 is used as the source domain data set (Market-1501- > DukeMTMC-reiD). The "base model" in the table represents the results obtained by model training without adding the reliable value sample sampling strategy and the new identity sample mining strategy, when training is performed using all training data in the source data domain and the target data domain. The "reliable samples" are the results obtained by model training after adding a reliable sample selection strategy, when training is performed using all training data in the source data domain and the selected reliable samples in the target data domain. The "reliable value samples" are the results obtained by model training after adding a reliable value sample selection strategy, when training is performed using all training data in the source data domain and the selected reliable and valuable samples in the target data domain. The "reliable value sample+new identity sample" is the result obtained by model training after adding the reliable value sample selection strategy and the new identity sample mining strategy, and at this time, training is performed using all training data in the source data domain, the reliable and valuable sample selected in the target data domain, and the new identity sample generated in the target domain.
From the comparison of the performances in table 1, it can be seen that the reliable value sample selection strategy and the new identity sample mining strategy are improved in both data sets compared with the performance of the basic model, and after the two strategies are combined, the model performance can be further improved obviously, so that the effectiveness of the invention is proved.
Table 1 comparison of reliable value sample selection policy and new identity sample mining policy for performance enhancement of pedestrian re-identification domain adaptation tasks
The foregoing description of the preferred embodiments of the invention is not intended to be limiting, but rather is intended to cover all modifications, equivalents, and alternatives falling within the spirit and principles of the invention.
Claims (5)
1. The pedestrian re-identification domain adaptation method based on reliable value sample and new identity sample mining is characterized by comprising the following steps:
step one: training a pedestrian target re-identification model F on the source domain data;
step two: clustering by using the model F as target domain data, and distributing pseudo tags according to a clustering result;
step three: selecting samples with high clustering reliability and high clustering value from the target domain data according to the clustering result, and forming a reliable and valuable sample set { X (X) in the target domain by corresponding pseudo labels r ,Y r -a }; the method is characterized by comprising the following substeps:
(3.1) for each sample in the target domain, calculating the clustering reliability according to the clustering result, and selecting samples with high reliability and corresponding pseudo tags to form a reliable sample set { X } c ,Y c -a }; the clustering reliability is a ratio of the similarity of the feature obtained by each target domain sample according to the model F and the closest clustering center to the feature, and the second closest clustering center, and the higher the ratio is, the higher the reliability is;
(3.2) for each sample in the target domain, calculating the clustering value according to the clustering result, and selecting samples with high value and corresponding pseudo tags to form a value sample set { X } v ,Y v -a }; the clustering value is dissimilarity between the characteristics obtained by each target domain sample according to the model F and the characteristics of other target domain samples, and the higher the dissimilarity is, the higher the value is; the other target domain sample features are kth other target domain sample features sequenced from high to low according to the feature similarity of the target domain sample and other target domain samples, and the dissimilarity is the reciprocal of the similarity;
(3.3) in the reliable sample set { X ] c ,Y c Sum value sample set { X } v ,Y v Finding co-occurring samples in the target domain, obtaining a reliable and valuable set of samples { X } r ,Y r }；
Step four: using source domain data { X ] s ,Y s Reliable and valuable sample set { X } in the target domain r ,Y r Generating a new identity sample set { X } n ,Y n }；
Step five: utilizing reliable and valuable sample sets { X in the target domain r ,Y r Sum of new identity sample set { X } n ,Y n Training an update model F;
step six: and repeating the second step to the fifth step until the model converges to obtain a final target domain model F.
2. The method of claim 1, wherein the clustering reliability and reliability sample set { X ] in (3.1) c ,Y c -calculated by:
(3.1.1) calculating a cluster center c of each cluster, which is obtained by averaging the characteristics of all samples in the clusters;
(3.1.2) for each target field sample { x } t ,y t Computing features F by model F t Find the nearest cluster center c i And the second closest cluster center c o ；
(3.1.3) calculating the target Domain sample { x } using t ,y t Cluster reliability of }:
wherein the method comprises the steps ofRepresenting the similarity between vector a and vector b;
(3.1.4) calculating the clustering reliability of all samples in the target domain, and averaging to obtain the mean clustering reliability tau g ；
(3.1.5) selecting samples with the clustering reliability greater than the mean clustering reliability, namely meeting R g ＞τ g Form a reliable sample set { X } c ,Y c }。
3. The method of claim 1, wherein the clustering value and value sample set { X ] in (3.2) v ,Y v -calculated by:
(3.2.1) for each target field sample { x } t ,y t Computing features F by model F t ；
(3.2.2) calculating the feature similarity of each target domain sample and other target domain samples and selecting the sample feature f of the k th bit of the ordered list according to the order from high to low k And obtain its dissimilarityAs the sampleClustering value R of the book l The method comprises the steps of carrying out a first treatment on the surface of the k is a positive integer;
(3.2.3) calculating the cluster value of all samples in the target domain and averaging to obtain the mean cluster value τ l ；
(3.2.4) selecting samples with cluster value greater than the average value cluster value, namely meeting R l ＞τ l Form a value sample set { X } v ,Y v }。
4. The method according to claim 1, wherein in the fourth step, a new identity sample set { X } n ,Y n -obtained by:
(4.1) randomly selecting one sample { x } in the source domain data s ,y s }；
(4.2) reliable and valuable sample set { X over target field r ,Y r Sequentially selecting one sample { x } r ,y r }；
(4.3) mixing the randomly selected samples and their labels in (4.1) and (4.2) respectively to obtain a new identity sample { x }, using the following formula n ,y n }：
x n ＝αx s +(1-α)x r
y n ＝Softmax(αy s +(1-α)y r )
Where α is a randomly sampled variable;
(4.4) repeating steps (4.1) to (4.3) for a reliable and valuable sample set { X over a target domain r ,Y r Each sample in the sequence generates a new identity sample to form a new identity sample set { X }, and n ,Y n }。
5. the method of claim 4, wherein in (4.3), α is randomly sampled from the distribution Beta (0.5 ).
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110935835.6A CN113657254B (en) | 2021-08-16 | 2021-08-16 | Pedestrian re-identification domain adaptation method based on reliable value sample and new identity sample mining |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110935835.6A CN113657254B (en) | 2021-08-16 | 2021-08-16 | Pedestrian re-identification domain adaptation method based on reliable value sample and new identity sample mining |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113657254A CN113657254A (en) | 2021-11-16 |
| CN113657254B true CN113657254B (en) | 2023-09-19 |
Family
ID=78479181
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110935835.6A Active CN113657254B (en) | 2021-08-16 | 2021-08-16 | Pedestrian re-identification domain adaptation method based on reliable value sample and new identity sample mining |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113657254B (en) |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109711366A (en) * | 2018-12-29 | 2019-05-03 | 浙江大学 | A kind of recognition methods again of the pedestrian based on group information loss function |
| WO2020186914A1 (en) * | 2019-03-20 | 2020-09-24 | 北京沃东天骏信息技术有限公司 | Person re-identification method and apparatus, and storage medium |
| CN111967294A (en) * | 2020-06-23 | 2020-11-20 | 南昌大学 | Unsupervised domain self-adaptive pedestrian re-identification method |
| CN111967325A (en) * | 2020-07-16 | 2020-11-20 | 长沙理工大学 | Unsupervised cross-domain pedestrian re-identification method based on incremental optimization |
| CN112016687A (en) * | 2020-08-20 | 2020-12-01 | 浙江大学 | Cross-domain pedestrian re-identification method based on complementary pseudo label |
| CN112633071A (en) * | 2020-11-30 | 2021-04-09 | 之江实验室 | Pedestrian re-identification data domain adaptation method based on data style decoupling content migration |
| CN112906606A (en) * | 2021-03-05 | 2021-06-04 | 南京航空航天大学 | Domain-adaptive pedestrian re-identification method based on mutual divergence learning |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160078359A1 (en) * | 2014-09-12 | 2016-03-17 | Xerox Corporation | System for domain adaptation with a domain-specific class means classifier |
-
2021
- 2021-08-16 CN CN202110935835.6A patent/CN113657254B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109711366A (en) * | 2018-12-29 | 2019-05-03 | 浙江大学 | A kind of recognition methods again of the pedestrian based on group information loss function |
| WO2020186914A1 (en) * | 2019-03-20 | 2020-09-24 | 北京沃东天骏信息技术有限公司 | Person re-identification method and apparatus, and storage medium |
| CN111967294A (en) * | 2020-06-23 | 2020-11-20 | 南昌大学 | Unsupervised domain self-adaptive pedestrian re-identification method |
| CN111967325A (en) * | 2020-07-16 | 2020-11-20 | 长沙理工大学 | Unsupervised cross-domain pedestrian re-identification method based on incremental optimization |
| CN112016687A (en) * | 2020-08-20 | 2020-12-01 | 浙江大学 | Cross-domain pedestrian re-identification method based on complementary pseudo label |
| CN112633071A (en) * | 2020-11-30 | 2021-04-09 | 之江实验室 | Pedestrian re-identification data domain adaptation method based on data style decoupling content migration |
| CN112906606A (en) * | 2021-03-05 | 2021-06-04 | 南京航空航天大学 | Domain-adaptive pedestrian re-identification method based on mutual divergence learning |
Non-Patent Citations (2)
| Title |
|---|
| 弱监督场景下的行人重识别研究综述;祁磊;于沛泽;高阳;;软件学报(第09期);全文 * |
| 运动场景下的时空域跟踪模型及原始-对偶算法;王诗言;于慧敏;;浙江大学学报(工学版)(第04期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113657254A (en) | 2021-11-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113378632B (en) | Pseudo-label optimization-based unsupervised domain adaptive pedestrian re-identification method | |
| Xu et al. | Deep learning for multiple object tracking: a survey | |
| Deng et al. | Variational prototype learning for deep face recognition | |
| CN109034448B (en) | Trajectory prediction method based on vehicle trajectory semantic analysis and deep belief network | |
| Lin et al. | Multi-domain adversarial feature generalization for person re-identification | |
| CN113326731B (en) | Cross-domain pedestrian re-identification method based on momentum network guidance | |
| CN111832511A (en) | Unsupervised pedestrian re-identification method for enhancing sample data | |
| Li et al. | Unsupervised person re-identification with multi-label learning guided self-paced clustering | |
| Zheng et al. | Soft pseudo-label shrinkage for unsupervised domain adaptive person re-identification | |
| Jiang et al. | Delving into sample loss curve to embrace noisy and imbalanced data | |
| CN112966647A (en) | Pedestrian re-identification method based on layer-by-layer clustering and enhanced discrimination | |
| CN112633071A (en) | Pedestrian re-identification data domain adaptation method based on data style decoupling content migration | |
| CN113140254A (en) | Meta-learning drug-target interaction prediction system and prediction method | |
| CN109117739A (en) | One kind identifying projection properties extracting method based on neighborhood sample orientation | |
| CN115841683A (en) | Light-weight pedestrian re-identification method combining multi-level features | |
| CN111695531A (en) | Cross-domain pedestrian re-identification method based on heterogeneous convolutional network | |
| Lin et al. | The Positive Effect of Attention Module in Few-Shot Learning for Plant Disease Recognition | |
| CN113837232A (en) | Black box model distillation method based on sample selection and weighting loss function | |
| CN113657254B (en) | Pedestrian re-identification domain adaptation method based on reliable value sample and new identity sample mining | |
| CN113611356A (en) | Drug relocation prediction method based on self-supervision graph representation learning | |
| Li et al. | FSRM-STS: Cross-dataset pedestrian retrieval based on a four-stage retrieval model with Selection–Translation–Selection | |
| CN116824270A (en) | Small sample vehicle type recognition method based on attention mechanism improvement | |
| Niemeijer et al. | Combining semantic self-supervision and self-training for domain adaptation in semantic segmentation | |
| CN113177521B (en) | Intelligent radiation source identification method based on combined twin network | |
| CN111814894B (en) | Multi-view semi-supervised classification method for rapid seed random walk |
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 |