A Three-Stage Framework with Reliable Sample Pool for Long-Tailed Classification

Synthetic Aperture Radar (SAR) imagery presents a promising solution for acquiring Earth surface information regardless of weather and daylight. However, the SAR dataset is commonly characterized by a long-tailed distribution due to the scarcity of samples from infrequent categories. In this work, we extend the problem to aerial view object classification in the SAR dataset with long-tailed distribution and a plethora of negative samples. Specifically, we propose a three-stage approach that employs a ResNet101 backbone for feature extraction, Class-balanced Focal Loss for class-level re-weighting, and reliable pseudo-labels generated through semi-supervised learning to improve model performance. Moreover, we introduce a Reliable Sample Pool (RSP) to enhance the model's confidence in predicting in-distribution data and mitigate the domain gap between the labeled and unlabeled sets. The proposed framework achieved a Top-1 Accuracy of 63.20% and an AUROC of 0.71 on the final dataset, winning the first place in track 1 of the PBVS 2023 Multi-modal Aerial View Object Classification Challenge.