Closed wshilton closed 1 year ago
Mask RCNN's success in object detection and segmentation resolves the use of the microwave sensor here. In particular, a trained M-RCNN can be applied on an optical data stream with a projection that is cued acoustically. M-RCNN also replaces the Haar cascade filter.
Since we are adopting DeepFace, we shall require M-RCNN to detect and segment only for human bodies for later body language analysis.
Since we are adopting DeepFace, we shall require M-RCNN to detect and segment only for human bodies for later body language analysis.
Also subject DeepFace processes only to those regions positively matched by M-RCNN.
Detectron2's DensePose performs the relevant detection, segmentation, mask generation, and pose estimation here.
Google's MediaPipe pose landmark detection is an alternative to be considered.
Pre-trained pose and hand landmark detection networks work sufficiently well to the extent that only integration work remains. Integration issues for this topic will now be raised on a case-by-case basis.
Ensure that the +-DOA regions identified by the acoustic and microwave sensors are of a size that leads to efficient detection by the human body Haar cascade filter.