Mask RCNN / Instance segmentation in PyTorch

I’m getting interested in PyTorch as an alternative to TF, for doing instance segmentation (via Mask RCNN or anything similar). Just starting to check into PyTorch, and learning the terrain.

Are there ‘standard’ PyTorch projects or code that is generally used as a base for Mask RCNN? Any docs on formats that are commonly used for training? IOW, the PyTorch equivalent of Tensorflow’s TFRecords. I’d like to find how to create and store the appropriate mask formats and such.


I think searching for “PyTorch detectron” on the web finds what you need. I’m not aware whether there is an official PyTorch implementation.

Best regards


That’s a great lead, Thomas. Most of the projects that I’ve seen are in an unspecified state, and I’d rather stay close to mainstream.

Now if there’s a way to parallel functionality of the Python interfaces with C++, I’ll be on track.

It seems like there should be a lot of fast C++ code for Mask RCNN and instance segmentation, given all the interest in self-driving cars. C++ would help reaction times.