CVPR：Refining Pseudo Labels with Clustering Consensus over Generations for Unsupervised Object Re-ID

Methods an overview

Improved unsupervised target recognition with pseudo tags based on generational clustering consensus .
1, In this paper, temporal embedding is introduced to regularize the noise pseudo tags in unsupervised target recognition .
2, In this paper, a pseudo label improvement strategy is proposed : In the iterative process of training, clustering consensus is used to improve the pseudo label .

Contents summary

Title of thesis	abbreviation	meeting / Periodical	Year of publication	baseline	backbone	Data sets
Refining Pseudo Labels With Clustering Consensus Over Generations for Unsupervised Object Re-Identification.	RLCC	CVPR	2021	【SpCL】Yixiao Ge, Feng Zhu, Dapeng Chen, Rui Zhao, and hong- sheng Li. Self-paced contrastive learning with hybrid mem- ory for domain adaptive object re-id. 7	ImageNet-pretrained [7] ResNet-50 [18] use DBSCAN [9] for clustering	Market-1501、DukeMTMC- reID、MSMT17

Online links ：https://openaccess.thecvf.com/content/CVPR2021/html/Zhang_Refining_Pseudo_Labels_With_Clustering_Consensus_Over_Generations_for_Unsupervised_CVPR_2021_paper.html
Source link ： -

Work Overview

1, we propose to properly estimate pseudo label similarities between consecutive training generations with clustering consensus and refine pseudo labels with tem- porally propagated and ensembled pseudo labels.
2, this is the first attempt to leverage the spirit of temporal ensembling [25] to improve classi- fication with dynamically changing classes over genera- tions.

Summary of results

With our proposed approach, state-of-the-art method [10] can be further boosted with up to 8.8% mAP improve- ments on the challenging MSMT17 [39] dataset.

Methods,

Methods the framework

Concrete realization

1,Y It means label ,M Represents a pseudo class . I Represents a sample set .
2,clustering consensus matrix Clustering consensus matrix C yes M（t-1） x M（t） Standard . The calculation is as follows: 1 Shown . among | · | Indicates the number of samples . Yes C Of normalization Handle It requires that the sum of the same column be 1 , The calculation is as follows 2 Shown .

3, There are two types of pseudo tag information ,1 yes hard Pseudo label 2 yes soft False tag confidence . spread hard The calculation of pseudo tags is as follows 3 Shown . spread soft The calculation of false label confidence is as follows 4 Shown . The calculation of the final pseudo tag is as follows 5 Shown .

4, The objective function of training is the formula 6.

experimental result

Overall evaluation

1, In fact, the innovation point is relatively simple .
2, From the picture 1 Look at ,HPLA You can also use this diagram . （ reflection , Why can people draw a picture that can send cvpr Paper graph of .）
3, Figure 1 and Figure 2 have a large amount of information repetition . Figure 2 is not a feeling framework.
4. The loss function doesn't seem very innovative , Just change a variable .

Small sample learning and intelligent frontier （ below ↓ official account ） The background to reply “RLCC", The paper electronic resources can be obtained .

Citation format

@inproceedings{DBLP:conf/cvpr/ZhangG0021,
author = {Xiao Zhang and
Yixiao Ge and
Yu Qiao and
Hongsheng Li},
title = {Refining Pseudo Labels With Clustering Consensus Over Generations
for Unsupervised Object Re-Identification},
booktitle = { {CVPR}},
pages = {3436–3445},
publisher = {Computer Vision Foundation / {IEEE}},
year = {2021}
}