Progressive Self-Guided Hardness Distillation for Fine-Grained Visual Classification

Fine-grained visual classification (FGVC) is a challenging problem due to its inherently high intra-class variances and low inter-class variances. Recently, vision transformers (ViTs) have demonstrated their powerful performance in both traditional and FGVC cases. However, compared with most categories, some categories are difficult to classify because of their similar characteristics and postures and their strong background interference, impacting the performance improvement achieved by the utilizd model. In this work, we present a novel method named progressive self-guided hardness distillation (PS-GHD), which defines a classification hardness judgement criterion and utilizes different approaches for the corresponding categories according to this criterion. This method gradually and correctly classifies various categories in three stages through knowledge distillation, so it can correctly classify indistinguishable categories to some extent. We demonstrate the value of PS-GHD by experimenting on four popular fine-grained benchmarks: CUB-200-2011, Nabirds, Stanford Cars, and Stanford Dogs. Our method achieves very competitive results on the four datasets. We also present qualitative results to enhance the interpretability of our model.