Skip to content
RetinaMask
Python Cuda C++ Shell
Branch: master
Clone or download
chengyangfu Update paths_catalog.py
update the http address of the pretrained model.
Latest commit 5bcf2e7 Mar 15, 2019
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
.github/ISSUE_TEMPLATE Add ISSUE_TEMPLATE for Github Oct 24, 2018
configs remove the unused sparsemask. Jan 14, 2019
demo Big commit. Nov 11, 2018
maskrcnn_benchmark Update paths_catalog.py Mar 15, 2019
tests
tools add log parser tool Nov 15, 2018
.flake8 Initial release Oct 24, 2018
.gitignore Add support for Python 2 (#11) Oct 26, 2018
ABSTRACTIONS.md Update ABSTRACTIONS.md (#82) Oct 31, 2018
CODE_OF_CONDUCT.md Initial release Oct 24, 2018
CONTRIBUTING.md Initial release Oct 24, 2018
INSTALL.md Add note about GCC < 4.9 (#38) Oct 26, 2018
LICENSE
MODEL_ZOO.md Initial release Oct 24, 2018
OLD_README.md Rename README.md to OLD_README.md Jan 5, 2019
README.md Update README.md Jan 31, 2019
TROUBLESHOOTING.md Add note about GCC < 4.9 (#38) Oct 26, 2018
arch.png Add files via upload Jan 7, 2019
run_test_R-50.sh Testing script. Nov 13, 2018
run_test_R-50_dist.sh Testing script. Nov 13, 2018
setup.py Add support for Python 2 (#11) Oct 26, 2018

README.md

RetinaMask

The code is based on the maskrcnn-benchmark.

alt text

Citing RetinaMask

Please cite RetinaMask in your publications if it helps your research:

@inproceedings{fu2019retinamask,
  title = {{RetinaMask}: Learning to predict masks improves state-of-the-art single-shot detection for free},
  author = {Fu, Cheng-Yang and  Shvets, Mykhailo and Berg, Alexander C.},
  booktitle = {arXiv preprint arXiv:1901.03353},
  year = {2019}
}

Contents

  1. Installation
  2. Models

Installation

Follow the maskrcnn-benchmark to install code and set up the dataset. Use config files in ./configs/retina/ for Training and Testing.

Models

Models BBox B(time) Mask M(time) Link
ResNet-50-FPN 39.4/58.6/42.3/21.9/42.0/51.0 0.124 34.9/55.7/37.1/15.1/36.7/50.4 0.139 link
ResNet-101-FPN 41.4/ 60.8/44.6/23.0/44.5/53.5 0.145 36.6/58.0/39.1/16.2/38.8/52.7 0.160 link
ResNet-101-FPN-GN 41.7/61.7/45.0/23.5/44.7/52.8 0.153 36.7/58.8/39.3/16.4/39.4/52.6 0.164 link
ResNeXt32x8d-101-FPN-GN 42.6/62.5/46.0/24.8/45.6/53.8 0.231 37.4/59.8/40.0/17.6/39.9/53.4 0.270 link

P.S. evaluation metric: AP, AP50, AP75, AP(small), AP(medium), AP(large), please refer to COCO for detailed explanation. The inference time is measured on Nvidia 1080Ti.

Run Inference

Use the following scripts. (Assume models are download to the ./models directory) Run Mask and BBox

python tools/test_net.py --config-file ./configs/retina/retinanet_mask_R-50-FPN_2x_adjust_std011_ms.yaml MODEL.WEIGHT ./models/retinanet_mask_R-50-FPN_2x_adjust_std011_ms_model.pth

Run BBox only

python tools/test_net.py --config-file ./configs/retina/retinanet_mask_R-50-FPN_2x_adjust_std011_ms.yaml MODEL.WEIGHT ./models/retinanet_mask_R-50-FPN_2x_adjust_std011_ms_model.pth MODEL.MASK_ON False

You can’t perform that action at this time.