altaykairat/ProxiBall-Testbench

ProxiBall Testbench: High-Speed Indoor Soccer Ball Detection

Dataset Summary

The ProxiBall Testbench is a Test subset of the main ProxiBall dataset (>18k images), used as a testbench in evaluation. Testbench set consists of unique annotated video frames other than in the main set, but strictly in the same environmental setup: indoor training arenas with artificial glass, artificial lightning, and close-proximity ball (2-10 meters from camera to ball). It is a highly curated, single-class benchmark dataset designed explicitly to evaluate object detection models on close-proximity, high-speed soccer balls. It serves as the official held-out evaluation split for the IEEE conference paper, "Close-Proximity High-Speed Soccer Ball Detection: Benchmarking and Data Augmentation Study."

Standard open-source sports datasets (e.g., SoccerNet, ISSIA-CNR) predominantly feature long-shot or broadcast-style views. In contrast, indoor training arenas (like Footbonaut or Footbot) feature compact scene geometry. When a soccer ball is struck at high speeds close to the camera, it undergoes extreme geometric deformation, resulting in elongated motion blur. This testbench isolates this specific domain gap, challenging models to maintain high precision and recall on severe edge cases where standard models usually fail.

• Paper: Close-Proximity High-Speed Soccer Ball Detection: Benchmarking and Data Augmentation Study (IEEE 2026)
• Authors: Altay Kairat, Azamat Shmitov, Yessimkhan Orynbay, and Arlen Smagulov
• Point of Contact: kairataltay@gmail.com

---

Dataset Structure

Data Instances

The dataset consists of 1,095 strictly held-out evaluation frames recorded at 1920x1080 resolution under artificial indoor lighting. To prevent data leakage, this testbench uses strictly unseen camera views (cam_south, cam_west) distinct from the proprietary training set.

• Total Images: 1,095
• Positive Images: 1,003 (contains a ball)
• Negative Images: 92 (hard backgrounds only, to test false-positive resistance)

Stratification (The "Ghost Ball" Problem)

To facilitate hard-case analysis, the positive frames in this testbench are naturally stratified across extreme variations:

• Size Buckets [normalized annotation area]:
  • Small [<0.00025]: 294 instances
  • Medium [<0.002]: 692 instances
  • Large [>0.002]: 17 instances
• Velocity/Blur Buckets [pixel ratio Width/Height]:
  • Slow [<1.05]: 226 instances
  • Medium [<1.3]: 630 instances
  • Fast [>1.3]: 147 instances (aspect ratios reaching up to 13.73 due to speed)

Data Fields and YOLO Format

Annotations are provided in standard YOLO .txt format. Each image has a corresponding .txt file with the same name.

• Format: class x_center y_center width height (normalized between 0 and 1)
• Class: 0 (soccer ball)

Example annotation line: 0 0.5432 0.8123 0.0345 0.0512

---

Dataset Creation

Curation Rationale

Collected from YouTube videos of Footbonaut/Footbot training sessions and recorded videos in enclosed indoor setups at Nazarbayev University Sports Center, Astana, Kazakhstan. The dataset was strictly manually filtered. Heavily occluded balls or context-poor frames were removed, while challenging but learnable blur, elongation, and near-edge cases were intentionally retained to test model robustness.

---

Usage

The files in the zip are ready to be directly dropped into a YOLO evaluation pipeline.

# Example data.yaml for YOLO testing
test: .../test/images
nc: 1
names: ['Ball']