# Roboflow Pronisi `golf-club-detection-1hgid` — project notes

**Pulled**: 2026-05-26 via Roboflow API (private key in `~/.config/roboflow/api_key`)
**Upstream**: https://universe.roboflow.com/pronisi/golf-club-detection-1hgid
**Version**: v3 (2023-06-26)
**License**: CC BY 4.0

## Scale

| Split | Images |
|---|---|
| train | 20 591 |
| valid | 1 277 |
| test  | 425 |
| **total** | **22 323** |

Original source: 8 577 images, 3× augmentation (hflip + ±45° rotation + ±15° shear) → 22 323.

## Format

COCO semantic-segmentation:
- `train/_annotations.coco.json` (and `valid/`, `test/`)
- Images 640×640 (stretch-resized), EXIF stripped
- 4 classes (`0`, `1`, `3`, `object`) with COCO-style polygon masks

## Class meaning

Labels are cryptic (`0`/`1`/`3`/`object`) — upstream didn't disclose a class-name mapping. From color hints in the API (`"0": #C7FC00`, `"1": #8622FF`, `"3": #FE0056`, `"object": #00FFCE`) and the project name, plausible interpretation:

- `0`/`1`/`3` likely correspond to shaft / head / grip (need to visually verify on a sample)
- `object` is a catch-all bbox-ish class

**TODO**: render a few samples to confirm class meaning before training.

## How this complements existing golf data

Pairs with **GolfDB swing clips**: train a club detector here, then run it across GolfDB frames to add per-frame club segmentation as a retrieval signal (where is the club at each phase). No other PD golf dataset offers club segmentation at this scale.