Balanced complement loss for long-tailed image classification

Long-tailed image classification is a significant research direction in the computer vision community, since many real-world image datasets exhibit an obvious long-tailed distribution. For a long-tailed dataset, head classes have most training samples, while tail classes occupy few training samples. In this scenario, samples of the head classes are treated as complement samples (all samples except the ground-truth class samples) of the tail classes, which generate overwhelmingly discouraging gradients on the tail classes. Consequently, the samples of tail classes are easily misclassified as the head classes. In view of this issue, we present a balanced complement (BACL) loss by introducing an adaptive weight for complement classes in the softmax cross-entropy loss. The adaptive weight helps to mitigate the overwhelming suppression of gradients from the complement samples for the tail classes, thereby balancing the gradient ratios between complement classes and ground-truth class. After that, we further propose a joint training framework by combining our method with normalized complement entropy (NCE) via a novel double-angle sine decay strategy. The proposed decay strategy is applied to adjust the contribution between the BACL and NCE losses at the different training stages. With our joint training framework, the model first learns useful information from the complement samples and then gradually turns its attention to the classification task. Experiments on long-tailed versions of CIFAR-10/100, SVHN and ImageNet-2012 datasets are conducted to reveal the significant effectiveness of the proposed methods.