A Weakly Supervised Part Detection Method for Robust Fine-Grained Classification

Fine-grained classification is a challenging task due to the subtle differences between subclasses, therefore, locating discriminative part features has become a key capability of fine-grained algorithms. In this paper, we design a sophisticated module to obtain the most discriminative part features by weakly supervised learning, which can be combined with multiple backbone networks, such as CNN-based or Transformer-based networks. Besides, we propose a self-knowledge distillation method that combines random homography transformation with a consistency loss in the training stage, which motivates the model to become more robust to pose variation and learns richer knowledge. The experimental results demonstrate that our model outperforms current fine-grained algorithms based on weakly supervised learning. Meanwhile, it follows a single stage design and does not require iterative optimization, thus having good training and inference efficiency.