wangsen1312 / unsupervised3dhuman

58 stars 5 forks source link

Evaluation script #14

Open ludvigsundoen opened 1 month ago

ludvigsundoen commented 1 month ago

Thanks for you interesting work. I am currently trying to recreate your evaluation on the CAPE dataset. I have generated point clouds from the validation set split of PTF (['00122', '00159', '00215']). Added gaussian noise of zero mean and 1mm standard-deviation. Then I subsampled the point cloud to 2048 points using farthest point sampling. Using your generate_pt.py script I find SMPL models and calculate the mean vertex-to-vertex error.

The results I get are 44.3 mm mean vertex-to-vertex error, compared to the 21.8 mm error that is reported in the paper. I also get 105.2 mm mean vertex-to-vertex error on the raw smpl output of the regressor.

Have I done something wrong? Or would it be possible to supply an evaluation script? Thanks!