Stop-Gradient Softmax Loss for Deep Metric Learning

Deep metric learning aims to learn a feature space that models the similarity between images, and normalization is critical step for boosting performance. However directly optimizing L2-normalized softmax loss cause network fail converge. Therefore some SOTA approaches appends scale layer after inner product relieve convergence problem, but it incurs new problem it's difficult best scaling parameters. In this letter, we look into characteristic of softmax-based propose novel objective function Stop-Gradient Softmax Loss (SGSL) solve in deep with L2-normalization. addition, found useful trick named Remove last BN-ReLU (RBR). It removes backbone reduce burden model. Experimental results on four fine-grained image retrieval benchmarks show our proposed approach outperforms most existing approaches, i.e., achieves 75.9% CUB-200-2011, 94.7% CARS196 83.1% SOP which other at least 1.7%, 2.9% 1.7% Recall@1.