A computer-vision and sports-analytics system for basketball. Muse consolidates the technical substrate for a set of products that aim to bridge the gap between the eye-test and counting stats -- for coaching, scouting, and training.
Basketball analysis has two traditions. The eye-test -- watching games to judge plays and players -- is rich but expensive and subjective. Counting stats -- points, assists, shooting percentages -- are cheap and comparable but lossy. Most of what happens in a possession never reaches the box score, yet that's where evaluations and coaching decisions get anchored.
Computer vision makes the eye-test computable. From raw game video, Muse extracts:
- Where the ball goes -- a continuous trajectory that makes possessions, passes, and shot attempts first-class events rather than summarised counts.
- How a shot is formed -- a jumpshot reduced to a vector of joint positions over time, directly comparable to reference-player templates.
Downstream products -- automated broadcasting for amateur leagues, shot-form coaching apps, play-by-play analytics, recruitment-ready game film -- consume those primitives.
| Subsystem | What it does | Core stack |
|---|---|---|
EyeBall/ |
Play detection and ball tracking: trajectory, passes, possession boundaries, play state | YOLO detection + Kalman filter |
FollowThrough/ |
Shooting-form pose analysis: jumpshot reduced to joint-vector template, compared against reference-player templates | MediaPipe Pose + Savitzky--Golay smoothing |
Each subsystem has its own README with design docs under <subsystem>/docs/.
The subsystems run independently today. Unifying their IO -- a shared typed trajectory/pose data product -- is on the roadmap.
- Python 3.12
- OpenCV 4.10, MediaPipe, NumPy, SciPy, Pandas, Shapely, imutils
requirements.txt at the repo root lists the top-level dependencies. The EyeBall YOLO path additionally requires a local PyTorch-YOLOv3 checkout -- see EyeBall/yolo_detection.md.
See each subsystem's README for its own entry points, methods, and status.
Muse/
├── EyeBall/ # Play detection and ball tracking
│ ├── assets/ # README figures
│ └── docs/ # Detection rationale, broadcasting application
├── FollowThrough/ # Pose / shot-form analysis
│ ├── source/ # Canonical implementation (OOP, MediaPipe Pose)
│ └── docs/ # Reserved for subsystem design docs
├── README.md
└── requirements.txt
Active development on main. Long-running solo project -- no release cadence, no external contributors.
Closed-source. All rights reserved. Copyright (c) 2026 Sameer Chaturvedi. Public visibility of this repository does not imply any grant of license. See LICENSE for terms, or contact sam.chaturvedi24@gmail.com for permission requests.