Combining Loss Reweighting and Sample Resampling for Long-Tailed Instance Segmentation

Traditional instance segmentation methods perform poorly when the training data has a long-tailed distribution. The recent long-tailed solutions only consider loss reweighting or sample resampling, which still suffers from the gradient imbalance of positive and negative samples and the overfitting risk of the tail classes. To address these problems, we propose a novel reweighting method, named Foreground and Background Separation Loss (FBSL), to alleviate the imbalance problem of the tail classes being suppressed by the overwhelming foreground and background during the learning process of the model. Moreover, we design a Feature Storage Module and Probabilistic Augmented Sampler (FSPAS) that performs targeted repetitive sampling based on the average probability of sample features, thereby alleviating the overfitting problem. With these two methods working together, the tail classes performance is significantly improved. Based on the Mask R-CNN framework, we achieve accuracies of 26.6% and 25.5% on LVIS v1.0 and COCO-LT datasets, respectively, exceeding the current mainstream methods.