Greetings Labelbox community. My team has been working on creating a custom pose dataset to improve Alphapose (and possibly OpenPose) keypoint detection. We’ve come up with a labeling approach that I feel is working quite well. We’re currently working on jobs that contain 36 labels (18 body keypoints and every keypoint has a visible and a not visible option). With the process we’ve come up with, our fast labelers take about 90 seconds to apply all keypoint labels to each identified person.
I’m posting because I’d love to meet and learn from others who are working on body poses. I imagine we could share ideas, techniques, problems encountered, quantitative/qualitative insights, etc. If interested, let me know here in the forum! And I’m open to ideas what the space we connect through looks like. It could be here, or it could routinely be over Zoom. If the latter I could start and demo what my team is doing internally.