This repository contains some extension works on the original maskrcnn benchmark repo at https://github.com/facebookresearch/maskrcnn-benchmark. You can pick the module you need and add them into the original maskrcnn repo for your research or product with no need of re-setup.
Below is a list of modules added.
- FCOS (https://arxiv.org/abs/1904.01355)
- cosine anealing learning step with warm up (https://arxiv.org/abs/1812.01187)
- squeeze and excitation net backbone (https://arxiv.org/abs/1709.01507)
- CBAM module backbone (https://arxiv.org/abs/1807.06521)
- LIBRA RCNN, with balanced IoU sampling, balanced feature map and balanced L1 loss implemented (https://arxiv.org/pdf/1904.02701.pdf)
- path aggregation neck (https://arxiv.org/pdf/1803.01534.pdf) PS: I don't add the xconv heavy head in this. The adaptive RoI pooling code doesn't work correctly in xconv. Only 2 MLP head with adaptive RoI pooling is implemented.
- cascade rcnn head (https://arxiv.org/abs/1712.00726)
- Global context block module in ResNet (https://arxiv.org/abs/1904.11492)
- hrnet backbone (https://arxiv.org/abs/1904.04514)
- negative sample training
- OHEM (in process)
If you have any questions, please start an issue. I will try my best to answer them. The training/inferencing config files should be more helpful with digging into codes. Upon request, I can upload some backbone models which are pre-trained on ImageNet.