Toward a More Robust Fine-Grained Fashion Retrieval

This paper aims to improve the robustness of fine-grained fashion retrieval (FGFR). The existing FGFR methods focus on designing complex attention modules for enhancing retrieval accuracy. They pay little attention to the characteristics of the real-world fashion data, which always contain item images and try-on images, and all images are taken from different viewpoints. This reduces the accuracy of existing FGFR models when applied to real-world applications. To solve this problem, we propose a novel contrastive learning method. Specifically, we first apply several geometric distortions (GD) to the input anchor image. Then, the distorted images are passed into existing FGFR models to learn an attribute-aware embedding. Finally, a contrastive loss is applied to ensure the distorted image is embedded close to the anchor image. The proposed method can be easily applied to existing FGFR models. Experiments on the FashionAI dataset verified its effectiveness in improving existing FGFR models by an average of approximately 5 percentage points over almost all the attributes. The proposed method also outperforms simply applying data augmentations to the input image by a significant margin. The source code is available in https://github.com/Dr-LingXiao/GD