Skip to content

Read You Like a "Mesh":Interactive Body Capture in First Person Video

Notifications You must be signed in to change notification settings

asetsuna/4DCapture-FPV

 
 

Repository files navigation

4D Human Body Capture from Egocentric Video via 3D Scene Grounding

[Project] [Paper]

Installation:

Our method requires the same dependencies as SMPLify-X and OpenPose. We refer to the official implementation fo SMPLify-X and OpenPose for installation details.

Our method also needs the installation of Chamfer Pytorch to calculate the chamfer distnace for enforceing human-scene constraints

Data Preparation:

Step 1: Dump video frames with desired fps (30) with utils/dump_videos.py. Run utils/split_frames to segment videos into equally long subatom clips. Repack frames to videos with utils/pack_videos.py (This is for faster openpose I/O).

Step 2: Run openpose_call.py under openpose folder to get human body keypoints, then run utils/openpose_helper to rename keypoint.json and run utils/openpose_filter.py to keep the most confident human keypoints.

Step 3: Run Smplify-X model with specified focal length and data directory. This step may take up to several hours. For instance:

python3 smplifyx/main.py --config cfg_files/fit_smplx.yaml  --data_folder /home/miao/data/rylm/downsampled_frames/miao_mainbuilding_0-1 --output_folder /home/miao/data/rylm/downsampled_frames/miao_mainbuilding_0-1/body_gen --visualize="False" --model_folder ./models --vposer_ckpt ./vposer --part_segm_fn smplx_parts_segm.pkl --focal_length 694.0

Step 4: Run Colmap for to generate scene mesh and camera trajectory. This step make take up to several hours depneding on the complexity of the scene. Then Run utils/camerpose_helper and utils/pointscloud_helper.py to generate desired points cloud file and camera pose.

Joint Optimization with 3D Scene Context:

Run global_optimization.py to conduct temproal smoothing and enforce human-scene constraints:

python3 global_optimization.py '/home/miao/data/rylm/packed_data/miao_mainbuidling_0-1/body_gen' '/home/miao/data/rylm/packed_data/miao_mainbuidling_0-1/smoothed_body

The resulting data should be organized as following:

  • datafolder:
    • videoname:
      • images: folder that contains all video frames
      • keypoints: folder that contains all body keypoints
      • body_gen: folder that contains all body mesh files:
      • smoothed_boyd: folder that contains all jointly-optimized body mesh files:
      • camera_pose.txt: text file that contains camera pose at each temporal footprint
      • meshed-poisson.ply: scene mesh file from dense reconstruction
      • camera.txt: text file that contains camera parameters
      • xyz.ply point cloud file. (use meash lab to convert .xyz file to .ply file)

Visualization in the World Coordinate:

Run global_vis.py to transform the body mesh in pivot coordinate to world coordinate. By default the viewpoint of open3d is the initial position camera trajectory. Setting bool flag to 'True' will resulting into a open3d viewpoint moving the same way as camera viewer.

python3 global_vis.py '/home/miao/data/rylm/downsampled_frames/miao_mainbuilding_0-1/' False

Visualization in the Egocentric Coordinate:

Run vis.py to view recosntrcuted body mesh on image plane.

python3 vis.py '/home/miao/data/rylm/segmented_data/miao_mainbuilding_0-1/'

Citation

If you find our code useful in your research, please use the following BibTeX entry for citation.

@inproceedings{liu20204d,
  title={4D Human Body Capture from Egocentric Video via 3D Scene Grounding},
  author={Liu, Miao and Yang, Dexin and Zhang, Yan and Cui, Zhaopeng and Rehg, James M and Tang, Siyu},
  booktitle={3DV},
  year={2021}
}

About

Read You Like a "Mesh":Interactive Body Capture in First Person Video

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%