search

zju3dv / mvpose

Code for "Fast and Robust Multi-Person 3D Pose Estimation from Multiple Views" (CVPR 2019, T-PAMI 2021)
https://zju3dv.github.io/mvpose/
Apache License 2.0
516 stars 79 forks source link

How to test on my own data? #23

Closed xiaobaozi1996 closed 5 years ago

xiaobaozi1996 commented 5 years ago

image I test it on my own data,the 2D pose estimation works well , but the 3D is wrong. I think it's because my project.mat is generated wrong. Could you please tell me the meanings of these paramters. I just use each camera's parameter to the correspondence in getK.m .Thank you in advance~

xiaobaozi1996 commented 5 years ago

My camera's paramters are as follows: image how should I set the right correapondence?

JiangWenPL commented 5 years ago

Hi Mao, those parameters are just in their literal meanings. For further information you could read the implementation in the mat2pickle.py. You can also have a look at our dataset if you have questions on how those parameters are used and transformed into fundamental matrixes. Feel free to ask us if you have other questions.

JulesDoe commented 5 years ago

This might get you in the right direction. I'm kind of shour there is still something wrong (PR if you find it ;). https://gist.github.com/JulesDoe/6b54602445d63f490709cecb4972a0f0

my results with this look like this: Screenshot from 2019-08-09 19-38-37

DongJT1996 commented 5 years ago

As you can see, the geometry affinity between person 0 and others is close to zero. You maybe can plot the epipolar lines corresponding to 2D keypoints of person 0 in other views.

JulesDoe commented 5 years ago

Hmm maybe ther went something wrong with the calibration of this particular kamera... thanks for the tip! I will look into it. Do i have to correct the distortion in the images bevor i pass them over?

JiangWenPL commented 5 years ago

We didn't un-distort the images in practice for the sake of convenience. But, theortically, it could be better if you do that.

lisa676 commented 4 years ago

@JulesDoe @JiangWenPL Hi, I've 6GB GPU (GTX1060). All the steps and compilations was done. When I try to run the Demo, it shows CUDA out of memory. When I changed the batch size from 32 to 4 in backend/tf_cpn/Detector2D.py (line 90) then it shows following error in Detector2D.py line 220. heatmaps = heatmaps[i], crops=crops[i] Kindly help me in this regard that how I can solve this problem? Thanks

sklf commented 4 years ago

Hello guys, I'm trying to use your code to estimate 3D poses. My dataset caontains frames from 12 cameras and camera parameters. I think my dataset is bigger than Campus and Shelf. So could you tell me what should I do to improve the performance? I just tried a little data. But the code seems very slow. Besides I can't get 3D pose plot at all. Could you help me?