Self-Supervised Modality-Aware Multiple Granularity Pre-Training for RGB-Infrared Person Re-Identification

RGB-Infrared person re-identification (RGB-IR ReID) aims to associate people across disjoint RGB and IR camera views. Currently, state-of-the-art performance of RGB-IR ReID is not as impressive that conventional ReID. Much due the notorious modality bias training issue brought by single-modality ImageNet pre-training, which might yield RGB-biased representations severely hinder cross-modality image retrieval. This paper makes first attempt tackle task from a pre-training perspective. We propose self-supervised solution, named Modality-Aware Multiple Granularity Learning (MMGL), directly trains models scratch only on multi-modal datasets, but achieving competitive results against without using any external data or sophisticated tuning tricks. First, we develop simple-but-effective ‘permutation recovery’ pretext globally maps shuffled images into shared latent permutation space, providing modality-invariant global for downstream tasks. Second, present part-aware cycle-contrastive (PCC) learning strategy utilizes cycle-consistency maximize agreement between semantically similar patches. enables contrastive unpaired scenarios, further improving discriminability local features laborious instance augmentation. Based these designs, MMGL effectively alleviates problem. Extensive experiments demonstrate it learns better (+8.03% Rank-1 accuracy) with faster speed (converge in few hours) higher efficiency (<5% size) than pre-training. The also suggest generalizes well various existing models, losses has promising transferability datasets. code will be released at https://github.com/hansonchen1996/MMGL.