sri299792458/quest-bowling-ball-yolo

Quest Bowling Ball YOLO Dataset

This dataset contains YOLO-format bounding-box annotations for detecting a bowling ball in Quest mixed-reality bowling footage.

The dataset was created for a live Quest-to-laptop bowling replay pipeline. YOLO is used only to find the first reliable ball seed; SAM2 then takes over for mask tracking and trajectory reconstruction.

Dataset Summary

• Task: single-class object detection
• Class: bowling_ball
• Format: YOLO detection labels
• Images: 801
• Positive images: 641
• Negative images: 160
• Source runs: 40 accepted bowling runs
• Capture domain: 2 bowling alleys
• Resolution: 960 x 720 source frames

Contributors

• Sri: https://github.com/sri299792458
• Apurv: https://github.com/ApurvK032
• Cale: https://github.com/CaleR2003

Project Repository

The full Quest bowling replay project is available here:

https://github.com/sri299792458/QuestBowlingStandalone

Splits

Split	Images	Positive	Negative	Source
Train	665	529	136	All Alley A + most Alley B
Val	71	59	12	Held-out Alley B
Test	65	53	12	Held-out Alley B

The validation and test sets are intentionally from the newer alley domain, so they measure cross-domain generalization rather than only in-domain memorization.

Directory Layout

images/
  train/
  val/
  test/
labels/
  train/
  val/
  test/
dataset.yaml
metadata/
  export_summary.json
  bowling_ball_combined_v2_split.json
training/
  YOLO26S_TRAINING_DECISIONS.md
  args.yaml
  results.csv

Each image has a matching .txt file in labels/<split>/. Negative examples have empty label files.

Usage With Ultralytics YOLO

yolo detect train model=yolo26s.pt data=dataset.yaml imgsz=1280

dataset.yaml:

path: .
train: images/train
val: images/val
test: images/test
names:
  0: bowling_ball

Annotation Workflow

This dataset was not hand-labeled frame by frame.

The workflow was:

1. Manually curate 40 bowling runs.
2. Select/verify ball seed frames.
3. Use SAM2-assisted tracking to propagate ball masks through each run.
4. Convert accepted SAM2 masks into YOLO bounding boxes.
5. Add pre-release negative frames so the detector learns when no ball is visible.
6. Split by bowling run, not by random image, to avoid leaking near-duplicate frames across train/val/test.

This gave a useful detector dataset from a small amount of manual labeling effort.

Intended Use

This dataset is intended for:

• bowling ball detection
• live sports replay prototypes
• small fast-object detection
• YOLO seed detection before a tracker such as SAM2
• AR/MR computer-vision experiments using Quest-style footage

It is not intended to be a general-purpose bowling dataset. It focuses on one object class and a specific camera/domain.

Limitations

• The dataset is small.
• Lighting and camera exposure vary by alley.
• The ball can be small, dark, blurred, or partially occluded.
• Labels are SAM2-assisted and review-filtered, not pure manual pixel inspection.
• The model trained from this dataset should be evaluated on new alleys before production use.

Project Context

This dataset was produced as part of the Quest Bowling Standalone project: a Quest 3 mixed-reality bowling replay system where YOLO detects the ball seed, SAM2 tracks the ball mask, and the laptop reconstructs a lane-space trajectory for headset replay.