Official PyTorch implementation of "I2L-MeshNet: Image-to-Lixel Prediction Network for Accurate 3D Human Pose and Mesh Estimation from a Single RGB Image", ECCV 2020
Could you please let me know why the regressor used for training uses the standard SMPL regressor + 5 keypoints, however the regressor for the datasets are quite different, for ex in COCO, pelvis is added as the mean of left and right hip, but in the original SMPL, the pelvis is above the hips, also the hips are way closer. Similarly for the HM36.
Wouldn't this inherently cause error in the 3D and 2D loss? or am I mistaken somewhere?
Could you please let me know why the regressor used for training uses the standard SMPL regressor + 5 keypoints, however the regressor for the datasets are quite different, for ex in COCO, pelvis is added as the mean of left and right hip, but in the original SMPL, the pelvis is above the hips, also the hips are way closer. Similarly for the HM36.
Wouldn't this inherently cause error in the 3D and 2D loss? or am I mistaken somewhere?