-
[pdf]
[bibtex]@InProceedings{Wang_2022_ACCV, author = {Wang, Kan and Hu, Shuping and Cheng, Jun and Cheng, Jun and Pang, Jianxin and Tan, Huan}, title = {RA Loss: Relation-Aware Loss for Robust Person Re-identification}, booktitle = {Proceedings of the Asian Conference on Computer Vision (ACCV)}, month = {December}, year = {2022}, pages = {177-194} }
RA Loss: Relation-Aware Loss for Robust Person Re-identification
Abstract
Previous relation-based losses in person re-identification (ReI-
D) typically comprise two sequential steps: they firstly sample both pos-
itive pair and negative pair and then deploy constraints to simultane-
ously improve intra-identity compactness and inter-identity separability.
However, existing relation-based losses usually place emphasis on ex-
ploring the relation between images and therefore consider only several
pairs during each optimization. This inevitably leads to different con-
vergence status for pairs of the same kind and brings about the intra-
pair variance problem. Accordingly, we propose a novel Relation-Aware
(RA) loss to address the intra-pair variance via exploring the informa-
tive relation across pairs. In brief, we introduce a macro-constraint and
a micro-constraint. The macro-constraint encourages the separation of
positive pair and negative pair via pushing far apart the two "center-
s" of the positive pair and the negative pair. The "center" of each kind
of pair are obtained via averaging all the pairs of the same kind. The
micro-constraint further enhances the compactness by minimizing the
discrepancies among pairs of the same kind. The two constraints work
cooperatively to relieve the intra-pair variance and improve the quali-
ty of pedestriansar representation. Results of extensive experiments on
three widely used ReID benchmarks, i.e., Market-1501, DukeMTMC-
ReID and CUHK03, demonstrate that the RA loss brings improvements
over existing relation-based losses.
Related Material