search

Arthur151 / ROMP

Monocular, One-stage, Regression of Multiple 3D People and their 3D positions & trajectories in camera & global coordinates. ROMP[ICCV21], BEV[CVPR22], TRACE[CVPR2023]
https://www.yusun.work/
Apache License 2.0
1.34k stars 229 forks source link

Poorly Segmented body #19

Open asimniazi63 opened 3 years ago

asimniazi63 commented 3 years ago

Poor results on some images, one of the examples is attached below:

images-fat1

How can we improve it

Arthur151 commented 3 years ago

Sorry for the failure in body shape estimation. The reason is that the body shape of the training data is not abundant. Most 3D pose dataset only have tens of subjects (actors). Therefore, the estimated shape tend to be normal weight. Most straight-forward solution is to fit the estimated body mesh to the body segmentation via a differentiable renderer (such as Example 2: Optimizing vertices of neural_renderer). The drawback is that you may have to take care of the number of iterative optimization. refPerformanceresults To avoid the bad results like above, it'd be better to optimize the estimated SMPL shape parameter and use its GMM prior constrain during loss calculation.