RAVEL-VB is a compact OpenVINO-powered model for detecting and tracking players and the ball in volleyball videos, with CPU inference and JSON/video outputs. It provides an inference script, model files, and configurations for temporal association and region-based detection.
Collecting history — the radar snapshots this repo daily. The trend line appears after 3 days of data (1 so far).
What it is
RAVEL-VB is a compact temporal detector for volleyball video. It detects players and the ball, associates player regions across frames, and can render an annotated demonstration video. This package contains a ready-to-run OpenVINO model and a standalone CPU inference script.
How it works
The architecture uses a 9-frame RGB clip input, 512 × 288 inference resolution. The player branch uses dynamic Top-K proposals, local multi-scale region sampling, persistent queries, and recurrent temporal association; the ball branch uses an independent temporal confidence-and-offset grid. It outputs player boxes with IDs and JSON predictions, and can render an annotated video.
Getting started
Install and run:
uv venv --python 3.12
uv pip install -r requirements.txt
uv run python infer_openvino.py input.mp4 \
--output out/predictions.json \
--output-video out/result.mp4
The bundled model is selected automatically. It can be overridden with
--model path/to/model.xml.
For a faster sparse pass, you can adjust stride and performance hints:
uv run python infer_openvino.py input.mp4 \
--stride 9 \
--performance-hint LATENCY \
--num-threads 4 \
--output out/predictions.json
Show all options:
uv run python infer_openvino.py --help
Recent releases
- none
Traction
13 stars
Behind the repo
Repository owner: asigatchov
Caveats
License: MIT Created: 2026-07-15 Last push: 2026-07-15 Language: Python Open issues: 0 Forks: 0