altaykairat/ProxiBall-Testbench

ProxiBall Testbench — cleaned 1,089-frame release

Version: clean-1089-20261015. Packaged on 15 October 2026; split corrected on 23 September 2026.

Overview

ProxiBall Testbench is the public evaluation subset of ProxiBall for single-class, close-proximity indoor soccer-ball detection, including small balls, motion blur, elongated appearances and empty background frames. It is a test-only release; the full training corpus and the separate Garage multi-camera dataset are not included.

The historical 1,095-frame testbench accompanied Close-Proximity High-Speed Soccer Ball Detection: Benchmarking and Data Augmentation Study. This release corrects that split; it is not a new dataset or a claim of a newly untouched test set.

Authors: Altay Kairat, Azamat Shmitov, Yessimkhan Orynbay, and Arlen Smagulov. Contact: kairataltay@gmail.com.

Verified contents

Item	Count / value
Images	1,089
YOLO label files	1,089
Positive images / ball boxes	997
Empty images / empty label files	92
Classes	1 (0: ball)
Distributed image resolution	960 × 960 pixels, RGB JPEG
Source-video filename groups	4

The files distributed here are 960 × 960, not 1920 × 1080. The old card's resolution description did not describe the actual distributed images. Images and labels are packaged byte-for-byte from the corrected local test split; no new resizing or relabelling is applied.

Source-video filename prefix	Images
IMG_0739_MOV	428
IMG_0740_MOV	415
Screencast-from-2026-03-13-15-47-18_mp4	43
testbench3_vid1_mp4	203

Split correction and evaluation scope

The previous release contained 1,095 images: 1,003 positive and 92 empty. Six positive frames from IMG_0738_MOV belonged to a source video also represented in training/validation. They were removed from test and reassigned to training in the full local dataset, leaving 1,089 test images: 997 positive and 92 empty. The six exact basenames are recorded in statistics.json and CHANGELOG.md.

The reference full dataset now contains 14,400 training, 2,538 validation and 1,089 test images. Only the test split is distributed in this repository. The remaining test source prefixes do not occur in local train/validation. Exact JPEG SHA-256 overlap with those splits was also checked and found absent. Train and validation still share source videos and should not be treated as source-video-independent splits.

These checks establish filename-prefix separation and absence of byte-identical JPEGs; they do not establish independence under every possible form of scene or near-duplicate overlap. Source-video identity is inferred from existing filenames, not independently reconstructed from original recordings. No unsupported claim of unseen camera views is made.

Historical scores on 1,095 images must remain labelled as such. Evaluate again on this release before reporting a score for 1,089 images. This public testbench has been evaluated repeatedly during research: freeze tuning choices on a separate validation set, disclose test reuse and do not describe it as a newly blind prospective evaluation.

Annotation format and diagnostics

Each JPEG has a same-stem text file containing either one YOLO row or no rows:

class_id x_center y_center width height

Coordinates are normalized to the distributed image dimensions; class_id is 0. Empty label files are intentional negative examples, not missing annotations. Annotation semantics follow the curated single-ball dataset; this is not a general multi-ball benchmark.

Normalized bounding-box area diagnostic: small <0.00025: 294; medium 0.00025 ≤ area < 0.002: 686; large area ≥ 0.002: 17.

Horizontal box width/height diagnostic: <1.05: 435; 1.05 ≤ ratio < 1.3: 430; ratio ≥ 1.3: 132. These bins describe annotation geometry, not ground-truth ball velocity. Width/height does not capture every blur direction; no speed, trajectory, 3D or synchronization labels are supplied. Values are recomputed from this release, not copied from the historical card.

Download and use

Download ProxiBall-Testbench-1089.zip, then extract it into a directory of your choice. Windows Explorer, macOS Archive Utility and standard ZIP tools can extract it.

data.yaml
README.md
CHANGELOG.md
statistics.json
manifest.jsonl
test/
  images/  # 1,089 JPEGs
  labels/  # 1,089 YOLO text files, including 92 empty files

data.yaml declares train: null; its val entry is an explicit compatibility alias for test/images, not an independent validation split. No training data are included. For an Ultralytics-based evaluator, the following Python example uses the test split explicitly; this is a usage example, not a bundled dependency or a new benchmark result:

from pathlib import Path
from ultralytics import YOLO

root = Path("/absolute/path/to/extracted/testbench")
YOLO("/absolute/path/to/your/model.pt").val(
    data=str(root / "data.yaml"),
    split="test", imgsz=960,
)

Do not train or tune hyperparameters using this test-only configuration. The val alias only accommodates evaluator schemas requiring that key. Record evaluator/version, checkpoint, input resolution, confidence/NMS settings and precision alongside AP50, AP50–95, precision and recall. Do not interpret unmeasured metrics or example settings as results.

manifest.jsonl records every image/label pair, source prefix, dimensions, object count and SHA-256 hashes. statistics.json contains counts, exclusions, audit checks and their limitations. SHA256SUMS beside the archive covers the release files, including the ZIP. The archive contains the same metadata except the external archive checksum file.

Provenance, rights and limitations

The parent ProxiBall corpus was curated from Footbonaut/Footbot material accessible through YouTube and team-recorded indoor footage. This is a parent-corpus provenance statement, not a verified per-frame mapping of this test subset to original video URLs or rights holders. Per-source original URLs and permissions have not been established by the packaging audit.

The existing repository metadata remains CC BY-NC 4.0; no new license grant over third-party source footage is asserted here. Public accessibility on YouTube does not itself establish redistribution permission. Third-party rights may apply independently; users should not infer clearance of original footage or commercial-use permission from this card. This release does not supply a rights audit or claim permission that has not been documented. Contact the maintainer about attribution, source identification or removal requests.

Frames from a small number of videos can be correlated. Domain coverage is limited to the curated indoor setting; results do not establish broad stadium, outdoor or multi-object generalization. Labels were manually curated, not independently re-annotated in this release.

Related materials and citation

Historical evaluation utilities: https://github.com/altaykairat/ProxiBall-dataset-eval-scripts (historical scripts may need the corrected split and evaluation configuration).

When using this release, cite the associated ProxiBall paper and explicitly identify ProxiBall-Testbench clean-1089-20261015, 1,089 images. Pin the Hugging Face commit revision and archive SHA-256 for reproducibility. No unverified DOI or publication metadata is supplied.