Contrastive dual-branch network for long-tailed visual recognition

Over the past decade, deep learning techniques have been widely applied to visual tasks, leading to remarkable breakthroughs. However, data in real-world scenarios often exhibit a long-tailed distribution, where head classes contain significantly more samples than tail classes. Models trained on such imbalanced data tend to bias the head classes, resulting in poor performance on tail classes. Moreover, due to the scarcity of samples in the tail classes, it is challenging for models to learn robust representations for those classes. Inspired by the success of contrastive learning in representation learning, we propose a Contrastive Dual-Branch Network (CDBN) for long-tailed visual recognition. CDBN integrates an imbalance learning branch and a contrastive learning branch to address the challenges of imbalanced data. The imbalance learning branch leverages traditional methods to address data imbalance, while the contrastive learning branch follows the principles of contrastive learning. Specifically, it uses two distinct data augmentation techniques to process the same batch of samples, generating positive sample pairs for enhanced learning. A contrastive auxiliary loss is then introduced to minimize the distance between these pairs in the normalized embedding space. Furthermore, we propose a Cumulative Fusion Strategy (CFS) to guide the model in progressively prioritizing tail classes throughout training. We conducted extensive experiments on the CIFAR10-LT, CIFAR100-LT, and ImageNet-LT datasets and compared our method with various advanced algorithms. The results demonstrate that our method substantially enhances performance across all datasets, achieving state-of-the-art results on several benchmarks. Our code is available at https://github.com/mmzbyxx/CDBN.