Video Analysis using vision models like Llama3.2 Vision and OpenAI's Whisper Models

A video analysis tool that combines vision models like Llama's 11B vision model and Whisper to create a description by taking key frames, feeding them to the vision model to get details. It uses the details from each frame and the transcript, if available, to describe what's happening in the video.

Features

• 💻 Can run completely locally - no cloud services or API keys needed
• ☁️ Or, leverage any OpenAI API compatible LLM service (openrouter, openai, etc) for speed and scale
• 🎬 Intelligent key frame extraction from videos
• 🔊 High-quality audio transcription using OpenAI's Whisper
• 👁️ Frame analysis using Ollama and Llama3.2 11B Vision Model
• 📝 Natural language descriptions of video content
• 🔄 Automatic handling of poor quality audio
• 📊 Detailed JSON output of analysis results
• ⚙️ Highly configurable through command line arguments or config file

Design

The system operates in three stages:

1. Frame Extraction & Audio Processing

   • Uses OpenCV to extract key frames
   • Processes audio using Whisper for transcription
   • Handles poor quality audio with confidence checks

2. Frame Analysis

   • Analyzes each frame using vision LLM
   • Each analysis includes context from previous frames
   • Maintains chronological progression
   • Uses frame_analysis.txt prompt template

3. Video Reconstruction

   • Combines frame analyses chronologically
   • Integrates audio transcript
   • Uses first frame to set the scene
   • Creates comprehensive video description

Requirements

System Requirements

• Python 3.11 or higher
• FFmpeg (required for audio processing)
• When running LLMs locally (not necessary when using openrouter)
  • At least 16GB RAM (32GB recommended)
  • GPU at least 12GB of VRAM or Apple M Series with at least 32GB

Installation

1. Clone the repository:

git clone https://github.com/byjlw/video-analyzer.git
cd video-analyzer

2. Create and activate a virtual environment:

python3 -m venv .venv
source .venv/bin/activate  # On Windows: .venv\Scripts\activate

3. Install the package:

pip install .  # For regular installation
# OR
pip install -e .  # For development installation

4. Install FFmpeg:

• Ubuntu/Debian:

  sudo apt-get update && sudo apt-get install -y ffmpeg

• macOS:

  brew install ffmpeg

• Windows:

  choco install ffmpeg

Ollama Setup

1. Install Ollama following the instructions at ollama.ai

2. Pull the default vision model:

ollama pull llama3.2-vision

3. Start the Ollama service:

ollama serve

OpenAI-compatible API Setup (Optional)

If you want to use OpenAI-compatible APIs (like OpenRouter or OpenAI) instead of Ollama:

1. Get an API key from your provider:

   • OpenRouter
   • OpenAI

2. Configure via command line:

   # For OpenRouter
   video-analyzer video.mp4 --client openai_api --api-key your-key --api-url https://openrouter.ai/api/v1 --model gpt-4o-mini
   
   # For OpenAI
   video-analyzer video.mp4 --client openai_api --api-key your-key --api-url https://api.openai.com/v1 --model meta-llama/llama-3.2-11b-vision-instruct

   Or add to config/config.json:

   {
     "clients": {
       "default": "openai_api",
       "openai_api": {
         "api_key": "your-api-key",
         "api_url": "https://openrouter.ai/api/v1"  # or https://api.openai.com/v1
       }
     }
   }

Note: With OpenRouter, you can use llama 3.2 11b vision for free by adding :free to the model name

Project Structure

video-analyzer/
├── config/
│   └── default_config.json
├── prompts/
│   └── frame_analysis/
│       ├── frame_analysis.txt
│       └── describe.txt
├── output/             # Generated during runtime
├── video_analyzer/     # Package source code
└── setup.py            # Package installation configuration

Usage

Basic Usage

Using Ollama (default):

video-analyzer path/to/video.mp4

Using OpenAI-compatible API:

video-analyzer path/to/video.mp4 --client openai_api --api-key your-key --api-url https://openrouter.ai/api/v1

Sample Output

Video Summary**\n\nDuration: 5 minutes and 67 seconds\n\nThe video begins with a person with long blonde hair, wearing a pink t-shirt and yellow shorts, standing in front of a black plastic tub or container on wheels. The ground appears to be covered in wood chips.\n\nAs the video progresses, the person remains facing away from the camera, looking down at something inside the tub. Their left hand is resting on their hip, while their right arm hangs loosely by their side. There are no new objects or people visible in this frame, but there appears to be some greenery and possibly fruit scattered around the ground behind the person.\n\nThe black plastic tub on wheels is present throughout the video, and the wood chips covering the ground remain consistent with those seen in Frame 0. The person's pink t-shirt matches the color of the shirt worn by the person in Frame 0.\n\nAs the video continues, the person remains stationary, looking down at something inside the tub. There are no significant changes or developments in this frame.\n\nThe key continuation point is to watch for the person to pick up an object from the tub and examine it more closely.\n\n**Key Continuation Points:**\n\n*   The person's pink t-shirt matches the color of the shirt worn by the person in Frame 0.\n*   The black plastic tub on wheels is also present in Frame 0.\n*   The wood chips covering the ground are consistent with those seen in Frame 0.

Advanced Usage

video-analyzer path/to/video.mp4 \
    --config custom_config.json \
    --output ./custom_output \
    --client openai_api \
    --api-key your-key \
    --api-url https://openrouter.ai/api/v1 \
    --model llama3.2-vision \
    --frames-per-minute 15 \
    --duration 60 \
    --whisper-model medium \
    --keep-frames

Command Line Arguments

Argument	Description	Default
video_path	Path to the input video file	(Required)
--config	Path to configuration directory	config/
--output	Output directory for analysis results	output/
--client	Client to use (ollama or openai_api)	ollama
--ollama-url	URL for the Ollama service	http://localhost:11434
--api-key	API key for OpenAI-compatible service	None
--api-url	API URL for OpenAI-compatible API	None
--model	Name of the vision model to use	llama3.2-vision
--frames-per-minute	Target number of frames to extract	10
--duration	Duration in seconds to process	None (full video)
--whisper-model	Whisper model size or model path	medium
--keep-frames	Keep extracted frames after analysis	False
--log-level	Set logging level (DEBUG, INFO, WARNING, ERROR, CRITICAL)	INFO
--language	Set language for transcription (if set as None, the language will be recognized)	None
--device	Select device to run Whisper model (cpu, cuda)	cpu

Configuration

The tool uses a cascading configuration system:

1. Command line arguments (highest priority)
2. User config (config/config.json)
3. Default config config/default_config.json

Configuration Options

General Settings

• clients.default: Default client to use (ollama or openai_api)
• clients.ollama.url: URL for the Ollama service
• clients.ollama.model: Vision model to use with Ollama
• clients.openai_api.api_key: API key for OpenAI-compatible service
• clients.openai_api.api_url: API URL for OpenAI-compatible API
• clients.openai_api.model: Vision model to use with OpenAI-compatible API
• prompt_dir: Directory containing prompt files
• output_dir: Directory for output files
• frames.per_minute: Target number of frames to extract per minute
• whisper_model: Whisper model size (tiny, base, small, medium, large) or Whisper model path. (For example, if using Windows, you can use E:\stt\models\models--Systran--faster-whisper-large-v3\snapshots\{UUID} to load your local model, or you can use relative path of folder {repo_path}\video_analyzer\video_analyzer)
• keep_frames: Whether to keep extracted frames after analysis
• prompt: Question to ask about the video

Frame Analysis Settings

• frames.analysis_threshold: Threshold for key frame detection
• frames.min_difference: Minimum difference between frames
• frames.max_count: Maximum number of frames to extract

Response Length Settings

• response_length.frame: Maximum length for frame analysis
• response_length.reconstruction: Maximum length for video reconstruction
• response_length.narrative: Maximum length for enhanced narrative

Audio Settings

• audio.sample_rate: Audio sample rate
• audio.channels: Number of audio channels
• audio.quality_threshold: Minimum quality threshold for transcription
• audio.chunk_length: Length of audio chunks for processing
• audio.language_confidence_threshold: Confidence threshold for language detection (the language will be detected in the first 30 seconds of audio.)
• audio.language: Set language for for transcription, default is None (If set, the language_confidence_threshold will not be used)

Output

The tool generates a JSON file (analysis.json) containing:

• Metadata about the analysis
• Audio transcript (if available)
• Frame-by-frame analysis
• Final video description

Example Output Structure

Uninstallation

To uninstall the package:

pip uninstall video-analyzer

License

MIT License

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.