gfl

Generalized Focal Loss Model(GFL)

Introduction

We reproduce the object detection results in the paper Generalized Focal Loss: Learning Qualified and Distributed Bounding Boxes for Dense Object Detection and Generalized Focal Loss V2. And We use a better performing pre-trained model and ResNet-vd structure to improve mAP.

Model Zoo

Backbone	Model	batch-size/GPU	lr schedule	FPS	Box AP	download	config
ResNet50	GFL	2	1x	----	41.0	model | log	config
ResNet50	GFL + CWD	2	2x	----	44.0	model | log	config1, config2
ResNet101-vd	GFL	2	2x	----	46.8	model | log	config
ResNet34-vd	GFL	2	1x	----	40.8	model | log	config
ResNet18-vd	GFL	2	1x	----	36.6	model | log	config
ResNet18-vd	GFL + LD	2	1x	----	38.2	model | log	config1, config2
ResNet50	GFLv2	2	1x	----	41.2	model | log	config

Notes:

• GFL is trained on COCO train2017 dataset with 8 GPUs and evaluated on val2017 results of mAP(IoU=0.5:0.95).

Citations

@article{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},
  journal={arXiv preprint arXiv:2006.04388},
  year={2020}
}

@article{li2020gflv2,
  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 arXiv:2011.12885},
  year={2020}
}