Maskrcnn-benchmark: Support for Fast RCNN

[adityaarun1]

❓ Questions and Help

Hi, Does this repository support Fast R-CNN?

I have a bunch of pre-computed bounding box proposals for each image, what changes are required to use these pre-computed proposals in the architecture?

[fmassa]

Fast R-CNN is not currently implemented, but here are the changes I see that should be made:

• add a proposals to https://github.com/facebookresearch/maskrcnn-benchmark/blob/d28845e112de36781b2b5f7217a34b2b62de8d2f/maskrcnn_benchmark/modeling/detector/generalized_rcnn.py#L33
• make the rpn be conditionally used https://github.com/facebookresearch/maskrcnn-benchmark/blob/d28845e112de36781b2b5f7217a34b2b62de8d2f/maskrcnn_benchmark/modeling/detector/generalized_rcnn.py#L50
• return the proposals from the dataset https://github.com/facebookresearch/maskrcnn-benchmark/blob/d28845e112de36781b2b5f7217a34b2b62de8d2f/maskrcnn_benchmark/data/datasets/coco.py#L59
• pass the proposals to the model for training https://github.com/facebookresearch/maskrcnn-benchmark/blob/d28845e112de36781b2b5f7217a34b2b62de8d2f/maskrcnn_benchmark/engine/trainer.py#L66
• and for inference https://github.com/facebookresearch/maskrcnn-benchmark/blob/d28845e112de36781b2b5f7217a34b2b62de8d2f/maskrcnn_benchmark/engine/inference.py#L24

That's I think pretty much it. There might be some other minimal changes.

Let me know how it goes.

[adityaarun1]

Thanks, this helps a lot. I will take a shot at implementing it.

[fmassa]

Cool, let us know when you manage to make it work!