Long-Tailed Instance Segmentation Using Gumbel Optimized Loss

Major advancements have been made in the field of object detection and segmentation recently. However, when it comes to rare categories, state-of-the-art methods fail detect them, resulting a significant performance gap between frequent categories. In this paper, we identify that Sigmoid or Softmax functions used deep detectors are major reason for low sub-optimal long-tailed segmentation. To address this, develop Gumbel Optimized Loss (GOL), It aligns with distribution classes imbalanced datasets, considering fact most expected probability. The proposed GOL significantly outperforms best method by $$1.1\%$$ on AP, boosts overall $$9.0\%$$ $$8.0\%$$ , particularly improving $$20.3\%$$ compared Mask-RCNN, LVIS dataset. Code available at: https://github.com/kostas1515/GOL .