batch_generation.md

Instructions & Remarks for 3D human joints batch generation

Precomputed bounding boxes

• Precomputed bboxes are needed to batch generate 3D skeletons from video inputs.

Basic structure of the output data

• Dictionary of numpy.ndarray
• Keys

1. vid_name: dim = [total_frame_number,]; name of the video that the frame belongs to, the frames are in temporal order, with fps=20.
2. joints3D: dim = [total_frame_number, 25, 3]; the joint hierachy is of format kinectv2; the name and the order of the "kinectv2" 25 joints can be found in "lib/data_utils/kp_utils.py"; please refer to joint_image.
3. bbox: dim = [total_frame_number, 4]; bounding boxes described in step#1, aligned with the number of frames (that are extracted by ffmpeg during 3D joint generation) of the video.

Run the generation

• Command.

python batch_generation.py --vid_folder 'folder containing the videos' --bbox_path 'path to to precomputed bbox' --pretrained_file 'path to the pretrained model'

• Parameters. The parameters specified at the begining of the code batch_generation.py can by customized.