Implementation of paper "Hand Pose Estimation via Latent 2.5 D Heatmap Regression" for 3D pose estimation of RGB hand pose dataset RHD
Preprocessed dataset has been uploaded to Google Drive at https://drive.google.com/file/d/1ahDGxYb6BmzQxRU_juv_QDFLbpHPMy29/view (~8GB)
Preprocessing VS Code is VisRHDAnnotation.cpp
See details about derivation of formula in ppt.
About caffe experiments:
-
RHDLatentHMMap
backbone network: ShuffleNet Model size: 4.33 MB root-relative 3D error: 25mm
-
RHDLatentHMMapRawPaper
backbone network: original conv-deconv network in paper Model size: 86.2 MB root-relative 3D error: 33mm
About caffe layers:
-
deep_hand_model_calc_norm_scale_layer
calculates normalized hand scale ("s" in raw paper)
-
deep_hand_model_solve_global_hand_scale_layer
predicts global hand scale during inference
-
deep_hand_model_solve_scale_normalized_global_location_layer
recovers scale normalized global joints from scale normalized root-relative 3d joints (just by adding wrist location)
-
deep_hand_model_solve_scale_normalized_global_z_root_layer
predicts global depth (z) of root: wrist joint
See usage in prototxt by searching keywords like "DeepHandModelSolveGlobalHandScale" "DeepHandModelSolveScaleNormalizedGlobalLocation" ...
@article{iqbal2018hand, title={Hand Pose Estimation via Latent 2.5 D Heatmap Regression}, author={Iqbal, Umar and Molchanov, Pavlo and Breuel, Thomas and Gall, Juergen and Kautz, Jan}, journal={arXiv preprint arXiv:1804.09534}, year={2018} }