Effect of Stage Training for Long-Tailed Multi-Label Image Classification

Yosuke Yamagishi, Shohei Hanaoka; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2023, pp. 2721-2728

In this study, we focus on the multi-stage training approach for training image classification models in the ICCV CVAMD 2023 Shared Task CXR-LT: Multi-Label Long-Tailed Classification on Chest X-Rays. In the proposed approach, the input image size and batch size are adjusted at each stage of the training process. In the first stage, we reduce the input image size and increase the batch size. Following that, we increase the image size and reduce the batch size in the second stage. A thorough search of the related literature did not yield validations of a similar approach for data with a long-tailed distribution. We successfully balance accelerated learning and performance by combining the proposed technique with various enhancements, such as oversampling, postprocessing using view positions, and ensemble methods, despite using a smaller model architecture and smaller input image size.