Computer Science > Computer Vision and Pattern Recognition
[Submitted on 7 Jul 2021 (v1), revised 28 Aug 2021 (this version, v2), latest version 16 Mar 2023 (v3)]
Title:Rethinking Sampling Strategies for Unsupervised Person Re-identification
View PDFAbstract:Unsupervised person re-identification (re-ID) remains a challenging task. While extensive research has focused on the framework design or loss function, we show in this paper that sampling strategy plays an equally important role. We analyze the reasons for differences in performance between various sampling strategies under the same framework and loss function. We suggest that deteriorated over-fitting is an important factor causing poor performance, and enhancing statistical stability can rectify this issue. Inspired by that, a simple yet effective approach is proposed, known as group sampling, which gathers groups of samples from the same class into a mini-batch. The model is thereby trained using normalized group samples, which helps to alleviate the effects associated with a single sample. Group sampling updates the pipeline of pseudo label generation by guaranteeing that samples are more efficiently divided into the correct classes. Group sampling regulates the representation learning process, which enhances statistical stability for feature representation in a progressive fashion. Qualitative and quantitative experiments on Market-1501, DukeMTMC-reID, and MSMT17 show that group sampling improves upon state-of-the-art methods by between 3.3%~6.1%. Code has been available at this https URL.
Submission history
From: Xumeng Han [view email][v1] Wed, 7 Jul 2021 05:39:58 UTC (2,738 KB)
[v2] Sat, 28 Aug 2021 04:26:41 UTC (8,116 KB)
[v3] Thu, 16 Mar 2023 07:43:08 UTC (2,602 KB)
References & Citations
Bibliographic and Citation Tools
Bibliographic Explorer (What is the Explorer?)
Litmaps (What is Litmaps?)
scite Smart Citations (What are Smart Citations?)
Code, Data and Media Associated with this Article
CatalyzeX Code Finder for Papers (What is CatalyzeX?)
DagsHub (What is DagsHub?)
Gotit.pub (What is GotitPub?)
Papers with Code (What is Papers with Code?)
ScienceCast (What is ScienceCast?)
Demos
Recommenders and Search Tools
Influence Flower (What are Influence Flowers?)
Connected Papers (What is Connected Papers?)
CORE Recommender (What is CORE?)
arXivLabs: experimental projects with community collaborators
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.