LAR IBRICS FORK of:

CameraHMR: Aligning People with Perspective (3DV 2025)

Original CameraHMR authors and pages

Priyanka Patel                Michael J. Black

🌐 Project Page | 📄 ArXiv Paper | 🎥 Video Results | 🌐 GitHub Repo

---

Figure: CameraHMR Results

---

🎬 Demo

Best way to use it is to use through https://github.com/upatras-lar/CameraHMR_environment/ that sets up a docker container with everything required to run in different modes

In case one wants to install without the above environment then use:

1. Create a conda environment and install all the requirements.

conda create -n camerahmr python=3.10
conda activate camerahmr
pip install torch==2.0.0 torchvision==0.15.1 torchaudio==2.0.1 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt

2. Download necessary files using:

Yolov8weights and Verify

wget https://github.com/ultralytics/assets/releases/download/v8.2.0/yolov8s.pt
md5sum yolov8s.pt
# 0a1d5d0619d5a24848a267ec1e3a3118  yolov8s.pt

3. Download models provided by camerahmr team

bash fetch_demo_data.sh

Alternatively, download files manually from the CameraHMR website. Ensure to update paths in constants.py if doing so manually.

Run the demo with following command. It will run demo on all images in the specified --image_folder, and save renderings of the reconstructions and the output mesh in --out_folder.

python demo.py --image_folder demo_images --output_folder output_images

IBRICS scripts

1. video_process.py

A simple utility that dumps all frames from a given video into a specified folder using ffmpeg for efficienty and speed.

Arguments

Option	Description	Default
video	Path to the input video file (e.g. input.mp4).	(required)
-o, --output-dir	Directory to save the PNG frames.	frames/

Example

python3 extract_frames.py my_video.mp4 -o extracted_frames

---

2. ibrics_main.py

Runs the HumanMeshEstimator on either a folder of pre‑extracted images or on a video (frames will be extracted first) and outputs a json file with either per-frame or per-joint structure.

Arguments

Option	Description	Default
--images <path>	Path to a folder of pre‑extracted image frames (mutually exclusive with --video).	(optional except if --video is missing)
--video <path>	Path to an input video file (frames will be extracted first).	(optional except if --images is missing)
-o, --output-dir	Directory to save extracted frames when using --video.	frames/
--separate	Flag to save the JSON file per joint instead of one per frame.	False
--json_name <name>	Base filename for output JSON (will be prefixed by per-frame_ or per-joint_).	/env/outputs/per-{frame or joint}_local_trajectories.json and /env/outputs/per-{frame or joint}_camera_trajectories.json

Note: You must supply exactly one of --images or --video.

Example: From Video

python3 ibrics_main.py --video my_video.mp4 -o extracted_frames --json_name results.json

# Extracts frames to ./extracted_frames/

# Runs mesh estimation on each frame

# Outputs SMPL trajectories to ./outputs/per-frame_results.json

Example: From Images

python3 ibrics_main.py --images extracted_frames --separate

# Reads frames from ./extracted_frames/

# Outputs the JSON per joint in ./outputs/per-joint_trajectories.json

Json Format

Save joint trajectories + SMPL parameters to JSON, with nested structure and shape info.

• per-frame mode:

{
    "frames": [
    { "trajectory": [[x,y,z],…], "parameters": {…} },
    … (one entry per frame)
    ],
    "average_parameters": {…},
    "shapes": {
    "num_frames": F,
    "joints_per_frame": J,
    "parameter_frames": F
    }
}

• per-joint mode:

{
    "trajectories": {
    "joint_0": [[…],[…],…],
    … (one entry per joint)
    },
    "parameters": {
    "frame_0": {…},
    "frame_1": {…},
    … (one entry per frame)
    },
    "average_parameters": {…},
    "shapes": {
    "num_joints": J,
    "frames_per_joint": F,
    "parameter_frames": F
    }
}

📚 Citation

If you find the original CameraHMR useful in your work, please cite:

@inproceedings{patel2024camerahmr,
title={{CameraHMR}: Aligning People with Perspective},
author={Patel, Priyanka and Black, Michael J.},
booktitle={International Conference on 3D Vision (3DV)},
year={2025} }