Intra-class Adaptive Augmentation with Neighbor Correction for Deep Metric Learning

Deep metric learning aims to learn an embedding space, where semantically similar samples are close together and dissimilar ones repelled against. To explore more hard informative training signals for augmentation generalization, recent methods focus on generating synthetic boost losses. However, these just use the deterministic class-independent generations (e.g., simple linear interpolation), which only can cover limited part of distribution spaces around original samples. They have overlooked wide characteristic changes different classes not model abundant intra-class variations generations. Therefore, generated lack rich semantics within certain class, but also might be noisy disturb training. In this paper, we propose a novel adaptive (IAA) framework deep learning. We reasonably estimate every class generate support mining Further, most datasets that few neighbor correction revise inaccurate estimations, according our correlation discovery generally variation distributions. Extensive experiments five benchmarks show method significantly improves outperforms state-of-the-art retrieval performances by 3%-6%. Our code is available at https://github.com/darkpromise98/IAA .