search

Nicholasli1995 / EvoSkeleton

Official project website for the CVPR 2020 paper (Oral Presentation) "Cascaded deep monocular 3D human pose estimation wth evolutionary training data"
https://arxiv.org/abs/2006.07778
MIT License
333 stars 44 forks source link

2D to 3D with own data #62

Closed dwygs closed 2 years ago

dwygs commented 2 years ago

hi I tried your 'inference.py' module with own 2d data, but the result has some problems： image image The reason is that your output need to use human3.6's 3d_mean and std to decode, but my data will produce errors if I use this decoding method, So I want to ask you how can I solve this kind of problem. thank you very much！

Nicholasli1995 commented 2 years ago

hi I tried your 'inference.py' module with own 2d data, but the result has some problems： image image The reason is that your output need to use human3.6's 3d_mean and std to decode, but my data will produce errors if I use this decoding method, So I want to ask you how can I solve this kind of problem. thank you very much！

Can you elaborate why the decoding method causes problem for your data?

dwygs commented 2 years ago

Thank you for your reply， The length of my bones is shorter than that of human 3.6, which is considered to be inclined to the depth direction, so I think the bone mean and variance calculated by human 3.6 are not applicable to my own data. Is this understanding correct? I want to know whether it is the cause of the network or the 'unNormalizeData' module . Thanks for your code and answers ！

Nicholasli1995 commented 2 years ago

Thank you for your reply， The length of my bones is shorter than that of human 3.6, which is considered to be inclined to the depth direction, so I think the bone mean and variance calculated by human 3.6 are not applicable to my own data. Is this understanding correct? I want to know whether it is the cause of the network or the 'unNormalizeData' module . Thanks for your code and answers ！

Your annotation for the leg maybe too short and the network thinks it corresponds to a person in another viewpoint. unNormalizeData should be OK but the input annotation style can be modified a bit. You may try lift the hip key-points higher to see if it helps. Otherwise you may also modify the training bone length/pose according to your need and re-train a network.

dwygs commented 2 years ago

Thanks！ I tried to modify the bone length and it can indeed be made correct.

sunmengnan commented 2 years ago

Thanks！ I tried to modify the bone length and it can indeed be made correct.

How did you find these length is shorter? as I see it, they are normal.

dwygs commented 2 years ago

How did you find these length is shorter? as I see it, they are normal.

You can see the picture I showed, the second one is rotated by an angle so that you can see the depth information, you can see that it is different from the pose in the image