Samples joint locations from given heatmaps and offset volumes

Criterion to compare the predicted per-joint heatmaps with the targets Underlying criterion can be either l1, l2, smooth l1, or per-voxel ce

Fixed learning rate after pre-defined number of warmup steps during which the learning rate linearly increases towards the final lr

Applying a Gaussian kernel to an n-dimensional tensor

Criterion to compare the predicted per-joint offset volumes with the target Underlying criterion can be either l1 (the default), l2, smooth l

Generate per-joint volumes where each voxel contains an offset to its skeleton joint Used to generate offset volumes for the input (and targe

Mean-Per-Joint-Position-Error (per-joint l2) between a predicted skeleton and a target skeleton, with pytorch

Positional encoder: Encodes the joint index prior to processing all joints with an attention module

Reduced skeleton, usually used for motion prediction. It removes all duplicate joints as well as hand and feet joints

The full skeleton, as recorded. Contains some duplicate joints: 0==11(hip), 13==16==24(Neck), 19==20(Left Hand), 22==23(Left Finger), 27==28(