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

By Xiang-Yang Li, Wenhai Wang, Lijun Wu, Shuo Chen, Xiaolin Hu, Jun Yu Li, Jinhui Tang, Jian Yang. 2020

1,924 citations recorded by the source corpus.

One-stage detector basically formulates object detection as dense classification and localization. The classification is usually optimized by Focal Loss and the box location is commonly learned under Dirac delta distribution. A recent trend for one-stage detectors is to introduce an individual prediction branch to estimate the quality of localization, where the predicted quality facilitates the classification to improve detection performance. This paper delves into the representations of the above three fundamental elements: quality estimation, classification and localization. Two problems are discovered in existing practices, including (1) the inconsistent usage of the quality estimation and classification between training and inference and (2) the inflexible Dirac delta distribution for localization when there is ambiguity and uncertainty in complex scenes. To address the problems, we design new representations for these elements. Specifically, we merge the quality estimation into the class prediction vector to form a joint representation of localization quality and classification, and use a vector to represent arbitrary distribution of box locations. The improved representations eliminate the inconsistency risk and accurately depict the flexible distribution in real data, but contain continuous labels, which is beyond the scope of Focal Loss. We then propose Generalized Focal Loss (GFL) that generalizes Focal Loss from its discrete form to the continuous version for successful optimization. On COCO test-dev, GFL achieves 45.0\% AP using ResNet-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 training settings. Notably, our best model can achieve a single-model single-scale AP of 48.2\%, at 10 FPS on a single 2080Ti GPU. Code and models are available at this https URL.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

Citations in the corpus

A Comprehensive Survey of Continual Learning: Theory, Method and Application (2023, methodology): “The idea of knowledge distillation is then introduced to other detection frameworks, such as CenterNet [508] (SID [328]), RetinaNet [259] (RILOD [247]), GFLV1 [250] (ERD [114]) and Faster R-CNN [358] (CIFRCN [152], Faster ILOD [327], DMC [493], BNC [96], IOD-ML [222]).”
PVT v2: Improved baselines with Pyramid Vision Transformer (2021, methodology): “We evaluate Swin Transformer and PVT v2 on four state-of-the-arts detectors, including Cascade R-CNN [1], ATSS [39], GFL [19], and Sparse R-CNN [28].”
Res2Net: A New Multi-Scale Backbone Architecture (2019, background): “, vectorized road extraction [63], object detection [35], weakly supervised semantic”
SDDGR: Stable Diffusion-Based Deep Generative Replay for Class Incremental Object Detection (2024, methodology): “1 shows that our method outperforms previous approaches such as LWF [33], RILOD [28], SID [43], and ERD [11] using GFLv1 [31], including CL-DETR [37].”
FCOS: A Simple and Strong Anchor-Free Object Detector (2020, positive): “built upon FCOS [38], [39], [40], [41].”
SuperYOLO: Super Resolution Assisted Object Detection in Multimodal Remote Sensing Imagery (2022, methodology): “…in this article, we selected 11 generic methods for comparison: one-stage algorithms ( YOLOv3 [47], FCOS [53], ATSS [54], RetainNet [51], and GFL [52]); two-stage method ( Faster R-CNN [5]); lightweight models ( MobileNetV2 [55] and ShufﬂeNet [56]); distillation-based methods ( ARSD [59]); and…”
YOLOv12: Attention-Centric Real-Time Object Detectors (2025, neutral): “Although improvements in YOLO have focused on areas such as loss functions [8, 35, 43, 44, 48, 67, 68], label assignment [22, 23, 34, 59, 69], network architecture design has remained a critical research priority [24, 28, 32, 57, 58].”
Real-Time Flying Object Detection with YOLOv8 (2023, background): “…includes the CIoU (complete IoU) loss proposed by Zheng et al.[22] as the box loss, the standard binary cross entropy for multi-label classiﬁcation as the classiﬁcation loss (allowing each cell to predict more than 1 class), and the distribution focal loss proposed by Li et al.[10] as the 3rd term.”
Class-Incremental Learning: A Survey (2023, background): “Moreover, the distillation can also be applied to other detectors ( e.g. , Faster R-CNN [256], GFL [257], RetinaNet [258], DERT [259] and CenterNet [260]) via intermediate features, which is similar to different distillation levels in CIL.”
A Review on Deep Learning in UAV Remote Sensing (2021, methodology): “The Libra-RCNN [147], ATSS [7], Guided Anchoring [61], FSAF [216], PAA [145], GFL [65], PISA [153] and VFNet [191] detectors explore di ff erent sampling strategies and new loss metrics to improve the quality of selected positive samples and reduce the weight of the large negative samples.”
Curriculum Temperature for Knowledge Distillation (2022, methodology): “Inspired by GFLv2 (Li et al. 2020a, 2021a), we propose to utilize the statistical information of probability distribution to control the smoothness of itself.”
UAV-YOLOv8: A Small-Object-Detection Model Based on Improved YOLOv8 for UAV Aerial Photography Scenarios (2023, methodology): “For the classiﬁcation task, binary cross-entropy loss (BCE Loss) is used predicted box bounding regression task, distribution Conventional FPN uses a top-down approach to convey deep semantic information. focal loss (DFL) [34] and CIoU employed.”