Generalized Focal Loss: Learning Qualified and Distributed Bounding Boxes for Dense Object Detection
See more comments in 大白话 Generalized Focal Loss(知乎)
[2020.11] GFocal has been adopted in NanoDet, a super efficient object detector on mobile devices, achieving same performance but 2x faster than YoLoV4-Tiny! More details are in YOLO之外的另一选择，手机端97FPS的Anchor-Free目标检测模型NanoDet现已开源~.
[2020.10] Good News! GFocal has been accepted in NeurIPs 2020 and GFocal V2 is on the way.
[2020.9] The winner (1st) of GigaVision (object detection and tracking) in ECCV 2020 workshop from DeepBlueAI team adopt GFocal in their solutions.
[2020.7] GFocal is officially included in MMDetection V2, many thanks to @ZwwWayne and @hellock for helping migrating the code.
One-stage detector basically formulates object detection as dense classification and localization (i.e., bounding box regression). 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 \emph{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 \emph{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 (i.e., separately trained but compositely used in test) and (2) the inflexible Dirac delta distribution for localization when there is ambiguity and uncertainty which is often the case 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 \emph{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 \emph{continuous} version for successful optimization. On COCO {\tt 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.
For details see GFocal. The speed-accuracy trade-off is as follows:
Please refer to INSTALL.md for installation and dataset preparation.
Please see GETTING_STARTED.md for the basic usage of MMDetection.
# assume that you are under the root directory of this project,
# and you have activated your virtual environment if needed.
# and with COCO dataset in 'data/coco/'
./tools/dist_train.sh configs/gfl_r50_1x.py 8 --validate./tools/dist_test.sh configs/gfl_r50_1x.py work_dirs/gfl_r50_1x/epoch_12.pth 8 --eval bboxCUDA_VISIBLE_DEVICES=0 python3 ./tools/benchmark.py configs/gfl_r50_1x.py work_dirs/gfl_r50_1x/epoch_12.pthFor your convenience, we provide the following trained models. All models are trained with 16 images in a mini-batch with 8 GPUs.
| Model | Multi-scale training | AP (minival) | AP (test-dev) | FPS | Link |
|---|---|---|---|---|---|
| GFL_R_50_FPN_1x | No | 40.2 | 40.3 | 19.4 | |
| GFL_R_50_FPN_2x | Yes | 42.8 | 43.1 | 19.4 | |
| GFL_R_101_FPN_2x | Yes | 44.9 | 45.0 | 14.6 | |
| GFL_dcnv2_R_101_FPN_2x | Yes | 47.2 | 47.3 | 12.7 | |
| GFL_X_101_32x4d_FPN_2x | Yes | 45.7 | 46.0 | 12.2 | |
| GFL_dcnv2_X_101_32x4d_FPN_2x | Yes | 48.3 | 48.2 | 10.0 |
[1] 1x and 2x mean the model is trained for 90K and 180K iterations, respectively.
[2] All results are obtained with a single model and without any test time data augmentation such as multi-scale, flipping and etc..
[3] dcnv2 denotes deformable convolutional networks v2. Note that for ResNe(X)t based models, we apply deformable convolutions from stage c3 to c5 in backbones.
[4] Refer to more details in config files in config/.
[5] FPS is tested with a single GeForce RTX 2080Ti GPU, using a batch size of 1.
Thanks MMDetection team for the wonderful open source project!
If you find GFL useful in your research, please consider citing this project.
@inproceedings{li2020generalized,
title={Generalized Focal Loss: Learning Qualified and Distributed Bounding Boxes for Dense Object Detection},
author={Li, Xiang and Wang, Wenhai and Wu, Lijun and Chen, Shuo and Hu, Xiaolin and Li, Jun and Tang, Jinhui and Yang, Jian},
booktitle={NeurIPS},
year={2020}
}
@article{li2020generalizedv2,
title={Generalized Focal Loss V2: Learning Reliable Localization Quality Estimation for Dense Object Detection},
author={Li, Xiang and Wang, Wenhai and Hu, Xiaolin and Li, Jun and Tang, Jinhui and Yang, Jian},
journal={arXiv preprint},
year={2020}
}
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent