Weakly-Supervised-3D-Hand-Pose-Estimation-from-Monocular-RGB-Images

Written by Liangjian Chen (Kupzhi@gmail.com)

Paper Reference:

ECCV18 weakly-supervised 3D hand

Preprocessing

Download STB dataset from here Unzip all the file into data/STB Run STB.py to get cropped hand

Download Pre-trained CPM model weight from here and put it into ./pretrained_weight

Training

Regression

Run Python train.py --cfg config/train/direct_regression.yaml to refined the pretrained_weight

Initial Depth Regularizer

find the best result in the previous training and put the path into the config/train/depth.yaml line 70 PRETRAINED_WEIGHT_PATH, and Run Python train.py --cfg config/train/direct_regression.yaml to initialized the weight of depth regularizer

End-to-end Training

find the best result in the previous training and put the path into the config/train/depth.yaml line 71 and 76 of PRETRAINED_WEIGHT_PATH, and Run Python train.py --cfg config/train/STB.yaml for the final end-to-end training