Closed timtensor closed 4 years ago
One thing I notice is that this implementation of hrnet applies the model to the cropped portion of the person returned from yolo whereas, from what I can tell, the model in the original paper is applied to the entire image. Losing the background context when predicting may affect the performance.
@bpeck81 thanks for the info , actually when i tried to disable yolov3
detector , it seemed to have worse performance, even for single person detection. I think it also has some limitation on multiperson detections , i am not sure .But with a set of optimized
åre trained wieghts , i could only manage detection of 2 persons with descent performance
Do you think , by changing the frame rate
we can improve the performance ?
According to the paper, HRNet should have quite higher performance than OpenPose when trained and tested on COCO. However, Openpose authors claim
In addition, our paper numbers are not based on the current models that have been released. We released our best model at the time but later found a better one.
therefore it may have better performance than HRNet. In my limited experience, performance of the two networks are similar.
@bpeck81 In the HRNet paper, authors state:
This paper is interested in single-person pose estimation
and
We extend the human detection box in height or width to a fixed aspect ratio: height:width = 4:3, and then crop the box from the image, which is resized to a fixed size, 256×192 or 384×288.
and
The two-stage top-down paradigm similar as [47, 11, 72] is used: detect the person instance using a person detector, and then predict detection keypoints.
Therefore, I add a YOLOv3 detector to find person instances and then analyze them with HRNet.
With the singleperson
option, the person detector is disabled and the image is directly analyzed by HRNet.
Thank you for answering the queries , As you said for multi person it does not work so good . Perhaps they will release new pretrained weights that would be better in performace. I tired on a multiperson
video , i was wondering how could be differentiate , which array corresponds to which person ?
For example if there are two person , the output array is of the type (2x17x2)
Then i wonder if there is an ID associated to each person , perhaps this is related as another question
At the moment, there is not an ID associated to each person because I didn't implement any person tracking functionality. Therefore, the order of the output is equal to the order of yolo detections.
@timtensor Could you please check the performance with the latest version of the code? I have implemented the idea proposed in #14 and, from my (limited) tests, accuracy is quite higher now in the multi-person setting.
@stefanopini i will try to test it in the coming days!
Yes i notice much better performance and quite stable as well
Hello I was playing around with different videos and clips to check performance of how pretrained models work for both Hrnet and Openpose. I seem to notice the open pose seem to have a better accuracy ? Have you tried it , what is your opinion about it ?