Learning How to Zoom In: Weakly Supervised ROI-Based-DAM for Fine-Grained Visual Classification

Wenjie Chen, Shuang Ran, Tian Wang, Lihong Cao

Abstract: Fine-grained visual classification (FGVC) is challenging due to the difficulty of finding discriminative features and insufficient labeled training data. How to efficiently localize the subtle but discriminative features with limited data is not straightforward. In this paper, we propose a simple yet efficient region of interest based data augmentation method (ROI-based-DAM) to handle the circumstance. The proposed ROI-based-DAM can first localize the most discriminative regions without the need of bounding box or part annotations. Based on these regions, ROI-based-DAM then carries out selective sampling and multi-scale cropping for constructing a series of high-quality ROI-based images. Thanks to its simplicity, our method can be easily implemented in the standard training and inference phases to boost the fined-grained classification accuracy. Our experimental results on extensive FGVC benchmark datasets show that the baseline model such as ResNeXt-50 can achieve competitive state-of-the-art performance by utilizing the proposed ROI-based-DAM, which demonstrate its effectiveness.

External IDs:dblp:conf/icann/ChenRWC21