Generalizable Person Re-identification Without Demographics
Generalizable Person Re-Identification (DG ReID) aims to learn ready-to-use cross-domain representations for direct cross-data evaluation. It typically fully exploit demographics information, e.g. the domain information and camera IDs to learn features that are domain-invariant. However, the protected demographic features are not often accessible due to privacy and regulation issues. Under this more realistic setting, distributionally robust optimization (DRO) provides a promising way for learning robust models that are able to perform well on a collection of possible data distributions (the ``uncertainty set”) without demographics. However, the convex condition of KL DRO may not hold for overparameterized neural networks, such that applying KL DRO often fails to generalize under distribution shifts in real scenarios. Instead, by applying the change-of-measure technique and the analytical solution of KL DRO, we propose a simple yet efficient approach, Unit DRO. Unit DRO minimizes the loss over a reweighted dataset where important samples (i.e. samples on which models perform poorly) will be upweighted and others will be downweighted. Empirical results show that Unit DRO achieves superior performance on large-scale DG ReID and cross-domain ReID benchmarks compared to standard baselines.
PDF Abstract