FreeAnchor

The Code for "FreeAnchor: Learning to Match Anchors for Visual Object Detection".

This repo is based on maskrcnn-benchmark, and FreeAnchor has also been implemented in mmdetection[link], thanks @yhcao6.

Detection performance on COCO:

Hardware	Backbone	Iteration	Scale jittering train / test	AP (minival)	AP (test-dev)	model link
4 x V100	ResNet-50-FPN	90k	N / N	38.6	39.1	Google Drive Baidu Drive
4 x V100	ResNet-101-FPN	90k	N / N	41.0	41.3	Google Drive Baidu Drive
4 x V100	ResNet-101-FPN	135k	N / N	41.3	41.8	Google Drive Baidu Drive
4 x V100	ResNeXt-101-32x8d-FPN	135k	Y / N	44.2	44.8	Google Drive Baidu Drive

Hardware	Backbone	Iteration	Scale jittering train / test	AP (minival)	AP (test-dev)	model link
8 x 2080Ti	ResNet-50-FPN	90k	N / N	38.4	38.9	Google Drive Baidu Drive
8 x 2080Ti	ResNet-101-FPN	90k	N / N	40.4	41.1	Google Drive Baidu Drive
8 x 2080Ti	ResNet-101-FPN	135k	N / N	41.1	41.5	Google Drive Baidu Drive
8 x 2080Ti	ResNeXt-101-32x8d-FPN	135k	Y / N	44.2	44.9	Google Drive Baidu Drive

Hardware	Backbone	Iteration	Scale jittering train / test	AP (minival)	AP (test-dev)	model link
8 x 2080Ti	ResNet-101-FPN	180k	Y / N	42.7	43.1	Google Drive Baidu Drive

Installation

Check INSTALL.md for installation instructions.

Usage

You will need to download the COCO dataset and configure your own paths to the datasets.

For that, all you need to do is to modify maskrcnn_benchmark/config/paths_catalog.py to point to the location where your dataset is stored.

Config Files

We provide four configuration files in the configs directory.

Backbone	Iteration	Scale jittering train / test	Config File
ResNet-50-FPN	90k	N / N	configs/free_anchor_R-50-FPN_1x.yaml
ResNet-101-FPN	90k	N / N	configs/free_anchor_R-101-FPN_1x.yaml
ResNet-101-FPN	135k	N / N	configs/free_anchor_R-101-FPN_1.5x.yaml
ResNeXt-101-32x8d-FPN	135k	Y / N	configs/free_anchor_X-101-FPN_j1.5x.yaml

Training with 4 GPUs (4 images per GPU)

cd path_to_free_anchor
export NGPUS=4
python -m torch.distributed.launch --nproc_per_node=$NGPUS tools/train_net.py --config-file "path/to/config/file.yaml"

Training with 8 GPUs (2 images per GPU)

cd path_to_free_anchor
export NGPUS=8
python -m torch.distributed.launch --nproc_per_node=$NGPUS tools/train_net.py --config-file "path/to/config/file.yaml"

Test on MS-COCO test-dev

cd path_to_free_anchor
python -m torch.distributed.launch --nproc_per_node=$NGPUS tools/test_net.py --config-file "path/to/config/file.yaml" MODEL.WEIGHT "path/to/.pth file" DATASETS.TEST "('coco_test-dev',)"

Evaluate NMS Recall

cd path_to_free_anchor
python  -m torch.distributed.launch --nproc_per_node=$NGPUS tools/eval_NR.py --config-file "path/to/config/file.yaml" MODEL.WEIGHT "path/to/.pth file"

Citations

Please consider citing our paper in your publications if the project helps your research.

@inproceedings{zhang2019freeanchor,
  title   =  {{FreeAnchor}: Learning to Match Anchors for Visual Object Detection},
  author  =  {Zhang, Xiaosong and Wan, Fang and Liu, Chang and Ji, Rongrong and Ye, Qixiang},
  booktitle =  {Neural Information Processing Systems},
  year    =  {2019}
}