quantumly/erebus-fixtures

Erebus Rescue Maze Fixtures

Image-classification dataset of RoboCup Junior Rescue Maze "fixtures" — the victim letters and hazardous-material placards a rescue robot must detect — captured from the robot's camera inside the Erebus simulation (Webots).

The images were hand-collected from the robot's camera in Erebus. This Rescue Maze category runs only in the simulator and has no physical counterpart, so these are the real images of the task — not a synthetic stand-in for anything. Used to train a small CNN fixture classifier.

Classes

Code	Meaning	Type
H	Harmed victim	victim letter
S	Stable victim	victim letter
U	Unharmed victim	victim letter
F	Flammable Gas	hazmat placard
C	Corrosive	hazmat placard
O	Organic Peroxide	hazmat placard
P	Poison	hazmat placard

Configs

Two variants are provided:

• simple (default) — 4 classes: the three victim letters (H, S, U) plus Flammable Gas (F). 7,452 images.
• complex — 7 classes: the victim letters plus all four hazmat placards (C, F, O, P). 4,405 images.

from datasets import load_dataset

ds = load_dataset("quantumly/erebus-fixtures")             # simple (default)
ds = load_dataset("quantumly/erebus-fixtures", "complex")  # full 7-class

Splits

Each config ships train / validation splits that reproduce the original training setup:

keras.utils.image_dataset_from_directory(validation_split=0.2, seed=1337)  # TF/Keras 2.12

i.e. classes sorted alphabetically, files sorted lexicographically within each class, the combined list shuffled with numpy.random.RandomState(1337), and the last 20 % taken as validation. The split is global, not class-stratified, so per-class validation ratios vary. splits/<config>.seed1337.csv lists every file's assignment.

Config	Train	Validation	Total
simple	5,962	1,490	7,452
complex	3,524	881	4,405

Per-class counts (train / validation)

simple — F 980 / 244 · H 1672 / 419 · S 1754 / 431 · U 1556 / 396

complex — C 418 / 96 · F 334 / 75 · H 569 / 153 · O 357 / 88 · P 380 / 105 · S 957 / 223 · U 509 / 141

Format

• 64×40 px PNG crops hand-collected from the robot's camera in the Erebus simulation.
• simple images are RGBA, complex are RGB; convert to RGB if you need uniformity across configs.

License

MIT.