README.md

---
title: Car Parts Image-to-Video Retrieval
emoji: 🚗
colorFrom: blue
colorTo: green
sdk: docker
pinned: false
license: mit
---

# Car Parts Image-to-Video Retrieval System

An intelligent system that detects car parts in images and retrieves matching video clips from an indexed automotive video.

## Features

- **YOLOv26s Detection**: Fine-tuned on car parts dataset
- **Semantic Matching**: Identifies doors, wheels, headlights, mirrors, bumpers, and more
- **Temporal Retrieval**: Returns precise video clip timestamps
- **Interactive Demo**: Upload any car image and find matching video segments

## How to Use

1. Upload an image containing car parts
2. The system detects all visible components
3. View matching video clips with timestamps
4. Each clip shows where that component appears in the source video

## Technical Details

- **Model**: YOLOv26s (small variant) fine-tuned for car part detection
- **Video Index**: Pre-computed detection index with bounding boxes and timestamps
- **Sampling Strategy**: Every 5th frame (4.8-6 FPS effective rate)
- **Clip Formation**: 3.0s gap threshold for temporal merging

## Assignment Context

This demo is part of **Assignment 2** for CS-UY 4613 Artificial Intelligence (Spring 2026).

**Student**: Hanze (James) Qiu  
**Repository**: [github.com/JamesQiu2005/CS-UY_4613_Assignments](https://github.com/JamesQiu2005/CS-UY_4613_Assignments)

---

Built with Ultralytics YOLO, OpenCV, and Gradio.