How does Mask R-CNN automatically output a different number of objects on the image?

Question

Recently, I was reading Pytorch's official tutorial about Mask R-CNN. When I run the code on colab, it turned out that it automatically outputs a different number of channels during prediction. If the image has 2 people on it, it would output a mask with the shape 2xHxW. If the image has 3 people on it, it would output the mask with the shape 3xHxW.

How does Mask R-CNN change the channels? Does it have a for loop inside it?

My guess is that it has region proposals and it outputs masks based on those regions, and then it thresholds them (it removes masks that have low probability prediction). Is this right?

Answer 1

Object detection models usually generate multiple detections per object. Duplicates are removed in a post-processing step called Non-Maximum Suppression (NMS).

The Pytorch code that performs this post-processing is called here in the RegionProposalNetwork class. The filtering loop you've mentioned performs the NMS and applies the score_thresh threshold (although it seems to be zero by default).