ProxyNCA++: Revisiting and Revitalizing Proxy Neighborhood Component Analysis

We consider the problem of distance metric learning (DML), where the task is\nto learn an effective similarity measure between images. We revisit ProxyNCA\nand incorporate several enhancements. We find that low temperature scaling is a\nperformance-critical component and explain why it works. Besides, we also\ndiscover that Global Max Pooling works better in general when compared to\nGlobal Average Pooling. Additionally, our proposed fast moving proxies also\naddresses small gradient issue of proxies, and this component synergizes well\nwith low temperature scaling and Global Max Pooling. Our enhanced model, called\nProxyNCA++, achieves a 22.9 percentage point average improvement of Recall@1\nacross four different zero-shot retrieval datasets compared to the original\nProxyNCA algorithm. Furthermore, we achieve state-of-the-art results on the\nCUB200, Cars196, Sop, and InShop datasets, achieving Recall@1 scores of 72.2,\n90.1, 81.4, and 90.9, respectively.\n