search

VideoObjectSearch / RandPerson

Surpassing Real-World Source Training Data: Random 3D Characters for Generalizable Person Re-Identification
https://arxiv.org/abs/2006.12774
Apache License 2.0
91 stars 18 forks source link

keypoints drift from their true positions #3

Open JasAva opened 4 years ago

JasAva commented 4 years ago

Thanks for the amazing dataset. I'm using Matlab to check the keypoints of randperson_subset (randperson_subset_point.txt). From my understanding, there are 9 pairs of (x,y), and start from the 3rd pair is the body keypoints, therefore I plot them on the corresponding images. However, I observed some keypoints are drifted from their true positions. Can you shed some insights on how to get the positions correct? Examples of drifts: imageimageimage

JasAva commented 4 years ago

Apart from the above question, I'm curious if the visualibity of the keypoints can be obtained (since there are occlusions in the images)?

Yanan-Wang-cs commented 4 years ago

Thank you for your attention. There is indeed a key point offset problem when the character moves, which may be because the character moves too fast or rendered too many characters. We currently have no rules for finding offset points, so we have not dealt with this problem. Maybe create a new video with slow movement and decrease the number of characters will help. As for the person-to-person occlusion of key points, it can be judged by the z value of the point coordinates, which is provided in the original data (video folder). Environmental occlusion can evaluate the occlusion by comparing the z-values ​​of people in different positions or record the z-values ​​of environmental occlusion objects in a new video.