Hi,
We've just open-sourced our PyTorch 1.0 implementation of Faster R-CNN and Mask R-CNN in https://github.com/facebookresearch/maskrcnn-benchmark.
It contains code for training and testing most of the detection and segmentation models from Detectron.
Compared to the current implementation in mmdetection, it seems to bring some significant speed-ups during training, with less memory usage as well.
Please feel free to try it out, and even plug in some of our components to potentially get some speedups.
@fmassa Congratulations to the release of the new implementation of mask rcnn! It is indeed a great piece of work. The team is now working intensely toward CVPR submission, and will learn from your efforts afterwards.
how about speed when infer?
We released our technical report for the latest benchmark.
Most helpful comment
@fmassa Congratulations to the release of the new implementation of mask rcnn! It is indeed a great piece of work. The team is now working intensely toward CVPR submission, and will learn from your efforts afterwards.