ATHENA Toolbox

ATHENA (Automatically Tracking Hands Expertly with No Annotations) is a Python-based toolbox designed to process multi-camera video recordings, extract 2D and 3D body and hand landmarks using MediaPipe, and perform triangulation and refinement of these landmarks. The toolbox provides a user-friendly GUI for selecting videos and configuring processing options.

Features

• Multi-Camera Processing: Handles multiple camera inputs for comprehensive analysis.
• MediaPipe Integration: Utilizes Google’s MediaPipe for extracting 2D landmarks of the body and hands.
• 3D Triangulation: Triangulates 2D landmarks from multiple cameras to reconstruct 3D landmarks.
• Smoothing and Refinement: Applies smoothing algorithms to refine the 3D landmarks.
• Parallel Processing: Supports multiprocessing to speed up the processing of large datasets.
• GUI for Easy Configuration: Provides a graphical user interface to select folders, recordings, and set processing options.
• Visualization: Offers options to save images and videos of the processed landmarks for visualization.

Installation

Prerequisites

• Operating System: Windows, macOS, or Linux.
• Python Version: Python 3.9 or higher.
• Hardware Requirements:
• CPU: Multi-core processor recommended for parallel processing.
• GPU (Optional): NVIDIA GPU for accelerated processing (if GPU processing is enabled).

Installation Steps

git clone https://github.com/neural-control-and-computation-lab/athena.git
cd athena
pip install .

It is recommended to create a conda environment before installing to manage dependencies and avoid conflicts.

Usage

1. Organize Your Videos

Place your video recordings in a main folder, structured as follows:

main_folder/
├── videos/
│   ├── recording1/
│   │   ├── cam0.avi
│   │   ├── cam1.avi
│   │   └── ...
│   ├── recording2/
│   │   ├── cam0.avi
│   │   ├── cam1.avi
│   │   └── ...
└── calibration/
    ├── cam0.yaml
    ├── cam1.yaml
    └── ...

• Each recording folder should contain video files from multiple cameras (e.g., cam0.avi, cam1.avi).
• The calibration folder should contain calibration files (.yaml) for each camera.

2. Ensure Calibration Files are Correct

• Calibration files should be labelled with camera names that match recorded videos.
• Calibration files should contain intrinsic and extrinsic parameters for each camera.
• The calibration files are essential for accurate triangulation of 3D landmarks.

3. Running the Toolbox

Launch the GUI:

athena

1. Select Main Folder and Recordings
   • Click on the “Select Folder” button to choose your main folder containing the videos and calibration directories.
   • A new window will appear, allowing you to select specific recordings to process. Select the desired recordings and click “Select”.
2. Configure Processing Options
   • General Settings:
     • Fraction of Frames to Process: Slide to select the fraction of frames you want to process (from 0 to 1). Default is 1.0 (process all frames).
     • Number of Parallel Processes: Choose the number of processes for parallel processing. Default is the number of available CPU cores.
     • GPU Processing: Check this option to enable GPU acceleration (requires compatible NVIDIA GPU).
   • MediaPipe Processing:
     • Run Mediapipe: Check this option to run MediaPipe for extracting 2D landmarks.
     • Save Images: Save images with landmarks drawn (can consume significant storage).
     • Save Video: Save videos with landmarks overlaid (requires “Save Images” to be enabled).
     • Minimum Hand Detection & Tracking Confidence: Adjust the confidence threshold for hand detection (default is 0.9).
     • Minimum Pose Detection & Tracking Confidence: Adjust the confidence threshold for pose detection (default is 0.9).
   • Triangulation:
     • Triangulation: Check this option to perform 3D triangulation of landmarks.
     • Save Images: Save images of the 3D landmarks.
     • Save Video: Save videos of the 3D landmarks.
3. Start Processing
   • Click the “GO” button to start processing.
   • A progress window will appear, showing the processing progress and average FPS.

4. Output Folder Structure

After processing, the toolbox will create additional directories within your main folder:

main_folder/
├── images/                # Contains images with landmarks drawn
├── imagesrefined/         # Contains refined images after triangulation
├── landmarks/             # Contains 2D and 3D landmark data
├── videos_processed/      # Contains processed videos with landmarks overlaid
└── ...                    # Original videos and calibration files

Troubleshooting

• MediaPipe Errors: Ensure that MediaPipe is correctly installed and compatible with your Python version. MediaPipe may have specific requirements, especially for GPU support.
• Calibration Mismatch: Verify that the number of calibration files matches the number of camera videos.
• High Memory Usage: Processing large videos or saving images and videos can consume significant memory and storage.
• Permission Issues: Ensure that you have read/write permissions for the directories where data is stored and processed.
• Conda Environment Activation: Always make sure that the Conda environment is activated before running the scripts.

Contributing

Contributions are welcome! If you encounter issues or have suggestions for improvements, please create an issue or submit a pull request on GitHub.

Frequently Asked Questions (FAQ)

Q: Do I need a GPU to run the ATHENA Toolbox?
A: No, a GPU is not required and does not generally speed up processing, which is already highly parallelized. If you have a compatible NVIDIA GPU, you can enable GPU processing in the options. Ensure that your GPU drivers and CUDA toolkit are correctly installed.

Q: Can I process videos from only one camera?
A: Yes, you can process videos from a single camera to extract 2D landmarks. However, triangulation to 3D landmarks requires synchronized videos from at least two cameras and their corresponding calibration files.

Q: How do I obtain the calibration files?
A: Calibration files are generated using camera calibration techniques, often involving capturing images of a known pattern (like a chessboard) from different angles. You can use OpenCV’s calibration tools or other software to create these .yaml files.

Q: The processing is slow. How can I speed it up?
A: You can increase the number of parallel processes if your CPU has more cores. Enabling GPU processing can also significantly speed up the landmark detection step. Additionally, processing a smaller fraction of frames can reduce computation time.

Q: Where can I find the output data after processing?
A: Processed data, images, and videos are saved in the images/, imagesrefined/, landmarks/, and videos_processed/ directories within your main folder.

Q: I encounter an error related to av or PyAV.
A: Ensure that av is installed via Conda using the conda install av -c conda-forge command, as it may have dependencies that are better handled by Conda.