Logo
← Back to projects

Roundnet Rally Segmentation

Team Full

A computer vision pipeline that turns raw roundnet (Spikeball) match footage into condensed highlight reels and live stats, automatically.

Project Manager
Henry Cheung

Computer Vision Sports Analytics
Ran
Fall 2026 – Spring 2027
Commitment
4-6 hrs/week
Openings
4-6

What you need coming in

Python required, plus some baseline exposure to computer vision and machine learning fundamentals — PyTorch basics, and familiarity with concepts like object detection (e.g. YOLO family models). OpenCV, multi-object tracking (ByteTrack, DeepSORT), and ffmpeg are learned on the job.

Overview

This project builds a computer vision pipeline that converts a full-length roundnet (Spikeball) match recording into a condensed highlight video of just the active rallies, generating live stats as a byproduct. The system detects rally boundaries and cuts dead time between points, tracks the ball and all players across frames, attributes each ball contact to a player, and classifies rally-ending events like aces, faults, and unforced errors.

Example of player segmentation

Roundnet is a fast-growing competitive sport with few existing automated video analysis tools, unlike volleyball, tennis, or basketball. The end goal is a tool a player or club could feed a raw phone/tripod recording into and get back a short, watchable highlight video plus a structured stat sheet — no manual editing or stat-keeping required.

Prerequisites

  • Python, with comfort using core ML libraries (PyTorch, NumPy, etc.)
  • Understanding of foundational ML/CV concepts (training a model, object detection, CNNs, etc.)
  • OpenCV, multi-object tracking (ByteTrack, DeepSORT), and/or ffmpeg are a plus

Milestones

Starts with a heuristic, motion/audio-based MVP for rally boundary detection built on collected and annotated match footage. From there, the project moves to trained ball and player detection/tracking models with court homography, then to contact attribution and event classification (aces, faults, unforced errors) replacing the earlier heuristic. The semester builds toward a full end-to-end pipeline demo: raw video in, condensed highlight video and stat sheet out.

Team structure

Members split into three subteams — Detection & Tracking, Segmentation & Event Logic, and Video/Infra & Demo.