2D Hand Keypoint Localization on Depth Video through Video to Video Translation
This repository contains the code for method introduced in: Markerless 2D Fingertip Localization on Depth Videos Using PairedVideo-To-Video Translation
To run the code you can, e.g., install the following requirements:
Please first download the example of test dataset for source and targert domain and put it under "datasets/handpose_5_new/".
Also, download pre-trained model from here and put it under "checkpoints/handpose_5_new/"
Next, compile a snapshot of FlowNet2 by running python util/download_flownet2.py
To get the translated sequenced from the first stage, run:
python stage1_test.py --name handpose_5_new --dataroot datasets/handpose_5_new/ --label_nc 0 --loadSize 128 --n_downsample_G 2 --use_real_img --how_many 8600
results will be generated and placed in "\results\handpose_5_new_test\test_latest\ [number] \"
Example of translated sequences from the first stage can be downloaded from here
Next, to perform the 2D keypoint localization in the second stage on example video result from the first stage, make sure the trasnlated sequenced are in the correct path as above before running stage2_test.py.
This code borrows heavily from vid2vid and pix2pixHD.