Fine-grained recognition via submodular optimization regulated progressive training

Progressive training has unfolded its superiority on a wide range of downstream tasks. However, it may fail in fine-grained recognition (FGR) due to special challenges with high intra-class and low inter-class variances. In this paper, we propose an active self-pace learning method to exploit the full potential of progressive training strategy in FGR. The key innovation of our design is to integrate submodular optimization and self-pace learning into a maximum-minimum optimization framework. The submodular optimization is regarded as a dynamic regularization to select active sample groups in each training round for restricting the search space of self-pace optimization. This can overcome the limitation of traditional self-pace learning that is easily trapped into local minimums when facing challenging samples. Extensive experiments on three public FGR datasets show that the proposed method can win at least 1.5% performance gain in various kinds of network backbones including swin-transformer.