search

microsoft / human-pose-estimation.pytorch

The project is an official implement of our ECCV2018 paper "Simple Baselines for Human Pose Estimation and Tracking(https://arxiv.org/abs/1804.06208)"
MIT License
2.93k stars 601 forks source link

how to do the NMS between bounding boxes from joints propagation and a detector? #103

Open aaab8b opened 5 years ago

aaab8b commented 5 years ago

I was trying to reproduce your algorithm of the tracking part. But there's a really specific detail that I can't figure out how to manage to complete. You got the bounding boxes from joints propagation and from the detector like R-FCN. However when you do the NMS operation, where do you get the confidence score of the bounding boxes from joints propagation?

lxy5513 commented 5 years ago

I also confused about, Could you please tell me the algorithm detail of bounding boxes from joints propagation and the implement of S_bbox, thanks in advance.

lxy5513 commented 5 years ago

Do you mean you generate multiple bounding boxes from joins propagation, the by one bbox from detector, NMS to generate a unified bbox ?

aaab8b commented 5 years ago

Do you mean you generate multiple bounding boxes from joins propagation, the by one bbox from detector, NMS to generate a unified bbox ? To be honest I have no idea how the author did this... But i tried to use a detector with two stage i.e. faster RCNN and concatenate the boxes from joints propagation and the proposals from RPN. Therefore i can get the confidence score from same domain

lxy5513 commented 5 years ago

concatenate the boxes from joints propagation and the proposals from RPN. Therefore i can get the confidence score from same domain

what's mean of concatenate the boxes?
I get predicted bbox (which by flownet2 and previous frame keypoints) and detection bbox(by yolov3 from this frame)
Then how can I do? how to use NMS, you know that predicted bbox is no confidence scores.

aaab8b commented 5 years ago

concatenate the boxes from joints propagation and the proposals from RPN. Therefore i can get the confidence score from same domain

what's mean of concatenate the boxes? I get predicted bbox (which by flownet2 and previous frame keypoints) and detection bbox(by yolov3 from this frame) Then how can I do? how to use NMS, you know that predicted bbox is no confidence scores. if you use one stage detector like YOLO, i don't know how to concatenate boxes. I use two stage detector so that the RPN will give proposals, and I simply treat the bboxes from flownet2S as another N proposals.

lxy5513 commented 5 years ago

Thank you a lot! you give a big inspiration