SPARE: Self-supervised part erasing for ultra-fine-grained visual categorization. (August 2022)

• Record Type: Journal Article
• Authors: Yu, Xiaohan; Zhao, Yang; Gao, Yongsheng
• Is Part Of: Pattern recognition. Volume 128(2022)
• Publication Date: 2022-08
• Subjects: Self-Supervised part erasing -- Ultra-fine-grained visual categorization -- Fine-grained visual categorization -- Random part erasing -- Weakly-supervised part segmentation; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2022.108691
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We propose a novel framework SPARE that segments parts using only image-level category labels and produces discriminative part feature representations for ultra-fine-grained visual categorization. A new self-supervised module is proposed to generate more diversified part segments with semantic meaning, and enhance the discriminability via predicting the contextual position of the erased parts. Very encouraging experimental results demonstrate the effectiveness of SPARE and the possibility in advancing research for ultra-fine-grained visual categorization. Abstract: This paper presents SPARE, a s elf-supervised par t e rasing framework for ultra-fine-grained visual categorization. The key insight of our model is to learn discriminative representations by encoding a self-supervised module that performs random part erasing and prediction on the contextual position of the erased parts. This drives the network to exploit intrinsic structure of data, i.e., understanding and recognizing the contextual information of the objects, thus facilitating more discriminative part-level representation. This also enhances the learning capability of the model by introducing more diversified training part segments with semantic meaning. We demonstrate that our approach is able to achieve strong performance on seven publicly available datasets covering ultra-fine-grained visual categorization and fine-grained visual categorization tasks.