In Mask-RCNN they modify the standard mask head for human pose keypoint detection with the following tweaks:
- Each keypoint is a 1-hot mask
- Instead of sigmoid non-linearity on the output of the final layer with spatial dim $m$x$m$, they use $m^2$-way softmax. This enforces the need to have a single keypoint in any given RoI.
I'm working with a task where I need to detect long string like objects and get their length- think of shoe lace loops or a pair of earphones laid out on the ground. It's very hard to properly annotate the data for a segmentation task as the image quality is not great. Moreover, I don't really need segmentation, I just need the skeleton. So I use an open-ended polyline for annotations.
So what sort of modifications would I make to the Mask-RCNN head? Some things I've considered:
- I could just dilate my polyline and it would be a mask. But it wouldn't be very faithful to the actual boundaries of the string like object, and so I'd be introducing a lot of label noise. (then again I don't need a really good IOU, I just need the length of the object to be right)
- I'm also worried about the aspect ratio. This might cause issues with bounding box regression. My anchors would need to have all sorts of ratios.