Generalized Focal Loss: Learning Qualified and Distributed Bounding Boxes for Dense Object Detection

One-stage detector basically formulates object detection as dense and localization. The classification is usually optimized by Loss and the box location is commonly learned under Dirac delta. A recent trend for one-stage detectors is to introduce an prediction branch to estimate the quality of localization, where the quality facilitates the classification to improve detection. This paper delves into the representations of the above three elements: quality estimation, classification and localization. Two are discovered in existing practices, including (1) the inconsistent of the quality estimation and classification between training and and (2) the inflexible Dirac delta distribution for localization when is ambiguity and uncertainty in complex scenes. To address the problems, design new representations for these elements. Specifically, we merge the estimation into the class prediction vector to form a joint of localization quality and classification, and use a vector to arbitrary distribution of box locations. The improved representations the inconsistency risk and accurately depict the flexible in real data, but contain continuous labels, which is beyond the of Focal Loss. We then propose Generalized Focal Loss (GFL) that Focal Loss from its discrete form to the continuous version for optimization. On COCO test-dev, GFL achieves 45. 0\\% AP using-101 backbone, surpassing state-of-the-art SAPD (43. 5\\%) and ATSS (43. 6\\%) with higher or comparable inference speed, under the same backbone and settings. Notably, our best model can achieve a single-model-scale AP of 48. 2\\%, at 10 FPS on a single 2080Ti GPU. Code and models available at https: //github. com/implus/GFocal.