Ysobel/SemDetect

NUpbr Ball Detection & Description Dataset

Synthetic renders of soccer balls (from the NUpbr generator) with bounding boxes and free-text visual descriptions, for training semantic ball detectors that generalise to ball appearances not seen during training.

Dataset structure

train/metadata.csv        # 1440 images, 72 ball instances
train/*.png
validation/metadata.csv   # 320 images, 16 ball instances
validation/*.png
test/metadata.csv         # 300 images, 15 ball instances
test/*.png
meta/*.yaml                # per-frame raw render annotation (provenance), one per image, matched by filename stem
ball_descriptions.csv      # one row per ball instance (103 rows): Ball Name, Description

Loads directly with:

from datasets import load_dataset
ds = load_dataset("imagefolder", data_dir=".")

metadata.csv columns

column	meaning
file_name	image filename, relative to the split folder
width, height	image dimensions in pixels
x1, y1, x2, y2	ball bounding box in pixels
class_name	always ball
instance_id	which of the 103 distinct ball instances appears in this image (e.g. ball_000)
description	free-text visual description of this ball instance, joined in from ball_descriptions.csv
meta_file	filename of the matching raw per-frame annotation in meta/

Split

The split is by ball instance, not by image: every image of a given ball instance lives entirely in one split. This means validation and test measure generalisation to ball appearances/descriptions the model never saw during training, not just held-out camera angles of a known ball.

split	images	instances
train	1440 (69.9%)	72
validation	320 (15.5%)	16
test	300 (14.6%)	15

Each ball instance contributes exactly 20 images across its assigned split.

Notes

• Filenames were renumbered sequentially (0000.png ... 2059.png) across the whole dataset.

License and acknowledgments

Released under the MIT license (see LICENSE). Images were rendered with NUpbr, a synthetic data generator built by the NUbots RoboCup team at the University of Newcastle, Australia; this dataset and the descriptions/annotations were produced independently by Ysobel Sims.