This project can be used for following appilcations:
The major challenge we ran into was the inavailablility of labelled dataset with occluded objects and their respective depth images. To solve this issue, we used the power of Computer Vision. On clear object with masked labels, we randomly drew rectangles of various colors accross the object thus creating an illusion of occulsion. A black (zeros) rectangle on the mask image at same position gave us the result of corresponding mask.
Another major issue we faced was result of Mask-RCNN led to multiple detection of same object. To overcome this issue, we introduced YOLOv2 network on top of Mask-RCNN network which led to good results for occluded objects.
Technologies used
Discussion