If you ever wanted to implement a Mask R-CNN from scratch in TensorFlow, you probably found Matterport’s implementation¹. This is a great one, if you only want to use a Mask R-CNN. However, as it is very robust and complex, it can be hard to thoroughly understand every bit of it. And the even bigger problem is, that it doesn’t run with new versions of TensorFlow.
And if you would have given a chance to a PyTorch implementation, the most frequently used one is the Detectron2², which is also very hard to understand because of its complexity.
From all the…
I’m a PhD student in Hungary at PPCU, in the field of Machine Learning. I am working at ULYSSYS Ltd. as a researcher in the field of diagnostics with AI.