MAL

MAL code for the paper Multiple Anchor Learning for Visual Object Detection pdf.

install

Get into MAL root folder.

1. Create conda env by conda env create -n MAL and activate it by 'conda activate MAL'.
2. Install python libraries. conda install ipython ninja yacs cython matplotlib tqdm
3. Install pytorch 1.1 + torchvision 0.2.1 by pip. download whl file at https://download.pytorch.org/whl/cu90/torch_stable.html pip install [downloaded file]
4. Install pycocotools pip install pycocotools
5. Copy https://github.com/facebookresearch/maskrcnn-benchmark/tree/master/maskrcnn_benchmark to this repository.
6. Build maskrcnn_benchmark by run python setup.py build develop
7. Install OpenCV3.

inference for an image

1. Go to ./demo
2. Run python image_demo.py. You can use your own image and change the image path in image_demo.py

test on COCO dataset

Get into MAL root folder. For test-dev set, run python python -m torch.distributed.launch --nproc_per_node=8 tools/test_net.py --config-file ./config/MAL_X-101-FPN_e2e.yaml MODEL.WEIGHT ./output/models/model_0180000.pth DATASETS.TEST "('coco_test-dev',)"

For val set, run python python -m torch.distributed.launch --nproc_per_node=8 tools/test_net.py --config-file ./config/MAL_X-101-FPN_e2e.yaml MODEL.WEIGHT ./output/models/model_0180000.pth

experimental result

mAP = 47.0 on test-dev

pre-trained model

ResNet50: https://share.weiyun.com/5kcZju5 ResNet101: https://share.weiyun.com/5gtr6Ho ResNext101: https://share.weiyun.com/oUZUWfSx