vnhngf/car-video-retrieval

Car Video Retrieval

You give it a picture of a car part (wheel, headlight, etc.) and it finds where that part shows up in a long car review video — with timestamps and YouTube links to jump straight to those moments.

How it works

• YOLOv8-World detects car parts in both the video and your query image (no training, just a custom list of part names).
• Detections are stored in Parquet so we can quickly look up “all frames where we saw a wheel,” etc.
• Retrieval runs the same detector on your query image, picks the top class, then returns the matching time ranges from the index.

What’s in this repo

• detections.parquet — Index of every detected object in the video: timestamp, class, bounding box, confidence.
• src/ — Scripts to download the video, extract frames, run detection, and run retrieval.
• The video itself isn’t included; the setup script downloads it (480p, 1 fps frames).

Detections schema

Column	Type	Meaning
video_id	string	e.g. rav4_review_2026
frame_index	int	Frame number (1 fps)
timestamp	float	Time in seconds
class_label	string	e.g. wheel, headlight
x_min, y_min, x_max, y_max	float	Bounding box
confidence_score	float	0–1

Run it

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

python src/01_setup_data.py   # download video, extract frames
python src/detect_video.py    # build the index
python src/03_retrieve.py     # run retrieval on the query set

Car-part vocabulary we use: bumper, door, fender, grille, door handle, headlight, hood, license plate, side mirror, roof, taillight, wheel, window, windshield, trunk, tire, rim.