mpi_inf_3dhp preprocess

nkolot / SPIN

Repository for the paper "Learning to Reconstruct 3D Human Pose and Shape via Model-fitting in the Loop"

Other

800 stars 170 forks source link

mpi_inf_3dhp preprocess #80

Open longbowzhang opened 4 years ago

longbowzhang commented 4 years ago

Hi, I just noticed that you use different annot3 dataset when you prepare the mpi_inf_3dhp train/test data. To be more specific, when dealing with the test data, you use univ_annot3 https://github.com/nkolot/SPIN/blob/277892a91e74af004c74457cd294bd102972f9b0/datasets/preprocess/mpi_inf_3dhp.py#L181 While for the train data, you use annot3 https://github.com/nkolot/SPIN/blob/277892a91e74af004c74457cd294bd102972f9b0/datasets/preprocess/mpi_inf_3dhp.py#L51 Is this a bug?

fabienbaradel commented 3 years ago

Hi @longbowzhang @geopavlakos, Any news about this issue? It seems to decrease a lot the performance on MPII_INF_3DHP. If we use 'annot3' keypoints at test time the performance is: MPJPE:118.78 - PA-MPJPE:73.09.