Learning from Noisy Anchors for
One-stage Object Detection

This repo contains the implementation for "Learning from Noisy Anchors for One-stage Object Detection" based on Detectron2.

Installation

See INSTALL.md. The following environment has been tested:

Python 3.7
CUDA 10.1
PyTorch 1.4.0
torchvision 0.5.0

Training and Testing

The config files are located at ./configs/COCO-Detection-NoisyAnchor. See GETTING_STARTED.md and follow the standard procedure to train/test RetinaNet with our method applied.

Before training, please download ImageNet pre-trained models as instructed in GETTING_STARTED.md and put them under ./outputs.

Results on COCO 2017 Val

RetinaNet:

Name	lr sched	box AP	download
R50-Baseline	1x	36.5	model
R50-NoisyAnchor	1x	38.6	model
R50-Baseline	3x	37.9	model
R50-NoisyAnchor	3x	40.2	model
R101-Baseline	3x	39.9	model
R101-NoisyAnchor	3x	42.0	model

Citation

If you find this project useful for your research, please use the following BibTeX entry:

@inproceedings{li2020learning,
  title={Learning from noisy anchors for one-stage object detection},
  author={Li, Hengduo and Wu, Zuxuan and Zhu, Chen and Xiong, Caiming and Socher, Richard and Davis, Larry S},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={10588--10597},
  year={2020}
}