How to use Mask-RCNN model to extract features of boxes?

vision
1

I’d like to use torchvision.models.detection.MaskRCNN model to detect objects and extract the corresponding features of bounding boxes of objects?

Anyone has any instruction or idea on how to do it properly?

2

I think you can use the bounding box coordinates and select the corresponding area of featuremap using those coordinates.

4

But if I use the bounding box coordinates and select the corresponding area of featuremap by using those coordinates, I would get features of different size for different boxes. What I want is features with same dimension.

5

I think you need to find a way to handle that. Maybe max pooling, average pooling, summation or sth like that. I’m not sure if resizing all of them to a fixed size would work or not.