Stella X. Yu : Papers / Google Scholar

Local Pseudo-Attributes for Long-Tailed Recognition
Dong-Jin Kim and Tsung-Wei Ke and Stella X. Yu
Neural Information Processing Systems Workshop on Self-Supervised Learning: Theory and Practice, New Orleans, Louisiana, 3 December 2022

We observe that solving the long-tailed distribution problems in real-world image datasets requires a fine-grained understanding of the local parts in an image. We propose a novel self-supervised learning framework with a new concept we call local pseudo-attribute (LPA) that is learned via clustering with local features without any extra human annotation. Note that the pseudo-attributes can be more balanced compared to the image-level class labels. Our final method using local pseudo-attributes achieves state-of-the-art performance through the experiments on various image classification setups with long-tailed distribution, such as CIFAR100- LT, iNaturalist, and ImageNet-LT datasets compared to the current long-tailed recognition methods.

attributes, local pseudo-attributes, long-tailed recognition