Boosting Fine-grained Fashion Retrieval with Relational Knowledge Distillation

Ling Xiao, Toshihiko Yamasaki; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 8229-8234

Fine-grained fashion retrieval (FGFR) aims to retrieve fashion items from a database that match specific and detailed attributes of a query image. This task requires a model to discern subtle variations which is more challenging than general recognition tasks. To improve retrieval accuracy we propose an online Knowledge Distillation (KD) framework that leverages KD's advantages in feature extraction. We also introduce a novel relational knowledge distillation (RKD) strategy that outperforms conventional KD by focusing on relational information. The proposed KD framework and RKD strategy can be easily applied to existing state-of-the-art FGFR models to significantly improve retrieval accuracy such as a +7.72% increase in mAP on the FashionAI dataset for ASENet_V2. The source code is available in https://github.com/Dr-LingXiao/RKD.