First of all thanks for the post! It helped me a lot in understanding the openpose code.

I was interested in your video of poses for a soccer game. I'm trying to do something similar and am struggling to get good results. When I run openpose on the video of the game, I only get poses for very few players that are closer to the camera. I tried detecting players first and than running openpose on each detection individually. That helps some but I still struggle to detect any poses for smaller players.

Do you have any tips, suggestions on how I can impove detection rate for smaller detections? I'm planning to try UltraHd resolution, but doubt that would help a lot.

Thanks again!!

Hi,
Sorry for the white colored text. It is an error in my theme. You can turn on the dark theme on the toolbar to see it.
Btw, I think that you should check the feature map to see if the small players are collapsed into a few pixels. If it is too small, then there is no way that the result is good. On quick approach is to divide the frame into multiple smaller & overlapped section and pass through OpenPose. However, this will slow down thing a lot. You can try AlphaPose to see if the result is better.

Ah, I see the night theme fixes it! :)

Thanks for such a quick response! Because of the size of some players I tried detecting every player in a frame using yolo frame detector. I end up with some player images that are 150x75 pixels in size. Than I run a player individually through openpose. That slows things down a lot, but for now I'm not too concerned about the speed.

When I run OpenPose detector on that one image, it resizes it to match the input shape of the network so it ends up processing it as a larger image. However I still only get results on images that have higher initial resolution. I assume becasue they are just more clear. Still I wonder if there is a way to improve it as I can still see clearly with my eyes the structure of the body in that small image.

Hi,
Based on your description, I can see that it is not how OpenPose is designed to do.
OpenPose follows the bottom-up approach where it detects the body parts then combine them into a pose. Therefore, it suppose to work with a full-size image with multiple people.
Your approach contains a person detector and pose detection so the input image is not similar with the training data of OpenPose.
In this case, I strongly suggest that you read the AlphaPose paper. They follows 2-step approach like you and they can also process in real-time.

Yeah I tried OpenPose with person detector because it seemed to work better. I just want to get the best results :)

Regarding your soccer video, what resolution did you use for the detection and did you run the entire frame through OpenPose or parts of it?

Not sure why my comment shows up as white colored text...