VLM-Video-Understanding

A minimalistic demo for image inference and video understanding using OpenCV, built on top of several popular open-source Vision-Language Models (VLMs). This repository provides Colab notebooks demonstrating how to apply these VLMs to video and image tasks using Python and Gradio.

Overview

This project showcases lightweight inference pipelines for the following:

• Video frame extraction and preprocessing
• Image-level inference with VLMs
• Real-time or pre-recorded video understanding
• OCR-based text extraction from video frames

Models Included

The repository supports a variety of open-source models and configurations, including:

• Aya-Vision-8B
• Florence-2-Base
• Gemma3-VL
• MiMo-VL-7B-RL
• MiMo-VL-7B-SFT
• Qwen2-VL
• Qwen2.5-VL
• Qwen-2VL-MessyOCR
• RolmOCR-Qwen2.5-VL
• olmOCR-Qwen2-VL
• typhoon-ocr-7b-Qwen2.5VL

Each model has a dedicated Colab notebook to help users understand how to use it with video inputs.

Technologies Used

• Python
• OpenCV – for video and image processing
• Gradio – for interactive UI
• Jupyter Notebooks – for easy experimentation
• Hugging Face Transformers – for loading VLMs

Folder Structure

├── Aya-Vision-8B/
├── Florence-2-Base/
├── Gemma3-VL/
├── MiMo-VL-7B-RL/
├── MiMo-VL-7B-SFT/
├── Qwen2-VL/
├── Qwen2.5-VL/
├── Qwen-2VL-MessyOCR/
├── RolmOCR-Qwen2.5-VL/
├── olmOCR-Qwen2-VL/
├── typhoon-ocr-7b-Qwen2.5VL/
├── LICENSE
└── README.md

Getting Started

1. Clone the repository:

git clone https://github.com/PRITHIVSAKTHIUR/VLM-Video-Understanding.git
cd VLM-Video-Understanding

2. Open any of the Colab notebooks and follow the instructions to run image or video inference.

3. Optionally, install dependencies locally:

pip install opencv-python gradio transformers

Hugging Face Dataset

The models and examples are supported by a dataset on Hugging Face:

VLM-Video-Understanding

License

This project is licensed under the Apache-2.0 License.