openmmla-vision 0.1.0.post2

Vision module for the OpenMMLA platform.

🎥 OpenMMLA Vision

Video module of the mBox - an open multimodal learning analytic platform. For more details, please refer to mBox System Design.

Table of Contents

• Related Modules
• Installation
  • Uber Server Setup
  • Video Base & Server Setup
  • Standalone Setup
• Usage
  • Realtime Indoor-Positioning
  • Video Frame Analyzer
• Visualization
• FAQ
• Citation
• References
• License

Related Modules

• mbox-uber
• mbox-audio

Installation

Uber Server Setup

Before setting up the video base, you need to set up a server hosting the InfluxDB, Redis, Mosquitto, and Nginx services. Please refer to mbox-uber module.

Video Base & Server Setup

1. Clone the repository

   git clone https://github.com/ucph-ccs/mbox-video.git
   

2. Install openmmla-vision

   Set up Conda environment

   # For Raspberry Pi
   wget "https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-$(uname)-$(uname -m).sh"
   bash Miniforge3-$(uname)-$(uname -m).sh
   
   # For Mac and Linux
   wget "https://repo.anaconda.com/miniconda/Miniconda3-latest-$(uname)-$(uname -m).sh"
   bash Miniconda3-latest-$(uname)-$(uname -m).sh
   

   Install Video Base

   conda create -c conda-forge -n video-base python=3.10.12 -y
   conda activate video-base
   pip install openmmla-vision[base]  # for Linux and Raspberry Pi
   pip install 'openmmla-vision[base]'  # for Mac
   

   Install Video Server

   The video server provides video frame analyzer services.

   conda create -c conda-forge -n video-server python=3.10.12 -y
   conda activate video-server
   pip install openmmla-vision[server]  # for Linux and Raspberry Pi
   pip install 'openmmla-vision[server]'  # for Mac
   

3. Set up folder structure

   cd mbox-video
   ./reset.sh
   

Standalone Setup

If you want to run the entire mBox Video system on a single machine, follow these steps:

1. Set up the Uber Server on your machine following the instructions in the mbox-uber module.

2. Install openmmla-vision with all dependencies:

   conda create -c conda-forge -n mbox-video python=3.10.12 -y
   conda activate mbox-video
   pip install openmmla-vision[all]  # for Linux and Raspberry Pi
   pip install 'openmmla-vision[all]'  # for Mac
   

3. Set up the folder structure:

   cd mbox-video
   ./reset.sh
   

This setup will allow you to run all components of mBox Video on a single machine.

Usage

Realtime Indoor-Positioning

1. Stream video from camera(s)

   • Distributed: stream on each camera host machine (e.g., Raspberry Pi)
   • Centralized: stream to a centralized RTMP server (e.g., MacBook, see Raspberry Pi RTMP streaming setup)

2. Calibrate camera's intrinsic parameters

   1. Print chessboard image from ./camera_calib/pattern/ and stick it on a flat surface
   2. Capture chessboard image with your camera and calibrate it by running ./calib_camera.sh

3. Synchronize multi-cameras' coordinate systems

   Calculate transformation matrix between main and alternative cameras:

   ./sync_camera.sh [-d <num_cameras>] [-s <num_sync_managers>]
   

   Default parameter settings:

   • -d: 2 (number of cameras to sync)
   • -s: 1 (number of camera sync manager)

   Modes:

   • Centralized:

     ./sync_camera.sh -d 2 -s 1
     

   • Distributed:

     # On camera host (e.g., Raspberry Pi)
     ./sync_camera.sh -d 1 -s 0
     # On synchronizer (e.g., MacBook)
     ./sync_camera.sh -d 0 -s 1
     

4. Run real-time indoor-positioning system

   ./run.sh [-b <num_bases>] [-s <num_synchronizers>] [-v <num_visualizers>] [-g <display_graphics>] [-r <record_frames>] [-v <store_visualizations>]
   

   Default parameter settings:

   • -b: 1 (number of video base)
   • -s: 1 (number of video base synchronizer)
   • -v: 1 (number of visualizer)
   • -g: true (display graphic window)
   • -r: false (record video frames as images)
   • -v: false (store real-time visualizations)

   Modes:

   • Centralized:

     ./run.sh
     

   • Distributed:

     # On camera host (e.g., Raspberry Pi)
     ./run.sh -b 1 -s 0 -v 0 -g false
     # On synchronizer (e.g., MacBook)
     ./run.sh -b 0 -s 1 -v 1
     

Video Frame Analyzer

1. Serve VLM and LLM on video server

   vllm

   vllm serve openbmb/MiniCPM-V-2_6 --dtype auto --max-model-len 2048 --port 8000 --api-key token-abc123 --gpu_memory_utilization 1 --trust-remote-code --enforce-eager
   vllm serve microsoft/Phi-3-small-128k-instruct --dtype auto --max-model-len 1028 --port 8001 --api-key token-abc123 --gpu_memory_utilization 0.8 --trust-remote-code --enforce-eager 
   

   ollama

   Install Ollama from official website.

   ollama pull llava:13b
   ollama pull llama3.1
   

2. Configure conf/video_base.ini

   [Server]
   backend = ollama
   top_p = 0.1
   temperature = 0
   vlm_model = llava:13b
   llm_model = llama3.1
   

3. Serve frame analyzer on video server

   cd examples/
   python serve_video_frame_analyzer.py
   

4. Run client script on video base

   python request_video_frame_analyzer.py
   

Visualization

After running the analyzers, logs and visualizations are stored in the /logs/ and /visualizations/ folders.

The following image shows a simple demo of the video frame analyzer:

FAQ

Citation

If you use this code in your research, please cite the following paper:

@inproceedings{inproceedings,
  author = {Li, Zaibei and Jensen, Martin and Nolte, Alexander and Spikol, Daniel},
  year = {2024},
  month = {03},
  pages = {785-791},
  title = {Field report for Platform mBox: Designing an Open MMLA Platform},
  doi = {10.1145/3636555.3636872}
}

References

• apriltags

License

This project is licensed under the MIT License - see the LICENSE file for details.