A dual-level part distillation network for fine-grained visual categorization

Fine-Grained Visual Categorization (FGVC) remains a formidable challenge due to large intra-class variation and small inter-class variation, which can only be recognized by local details. Existing methods adopt part detection modules to localize discriminative regions for extracting part-level features, which offer crucial supplementary information for FGVC. However, these methods suffer from high computational complexity stemming from part detection and part-level feature extraction, while also lacking connectivity between different parts. To solve these problems, we propose a Dual-level Part Distillation Network (DPD-Net) for FGVC. Our DPD-Net extracts features at both object and part levels. In the object level, we first use residual networks to extract middle and high level features for generating middle and high object-level predictions, and concatenate these two predictions to produce the final output. In the part level, we use a part detection module to localize discriminative parts for extracting part-level features, point-wisely add features of different parts to generate an averaged part-level prediction, and concatenate different part features to produce a concatenated part-level prediction. We use knowledge distillation to transfer information from the averaged and concatenated part-level predictions to the middle and high object-level predictions, respectively. To supervise the training of our method, we design five losses, namely the pair-wise consistency of detected parts, the concatenated final prediction, the averaged part-level prediction, the cosine-embedding loss, and the concatenated part-level prediction. Experimental results show that our DPD-Net achieves state-of-the-art performance on three Fine-Grained Visual Recognition benchmarks. In addition, our DPD-Net can be trained end-to-end without extra annotations.