CRUMB: the Collected Radiogalaxies Using MiraBest dataset

The CRUMB dataset is a machine learning image dataset of Fanaroff-Riley galaxies, constructed by combining the datasets of MiraBest, FR-DEEP, AT17 and a supplementary MiraBest hybrid dataset.

Sources are labelled using a unified "basic" label system with the following classes:

0: FRI

1: FRII

2: Hybrid source

The original labels of each of the parent datasets are also retained as a "complete" label which is accessible using the built-in "complete_labels" method on the dataloader. These labels are as follows:

MiraBest: 0 (confidently-classified FRI), 1 (confidently-classified wide-angle-tail), 2 (confidently-classified head-tail), 3 (uncertainly-classified FRI), 4 (uncertainly-classified wide-angle-tail), 5 (confidently-classified FRII), 6 (confidently-classified double-double), 7 (uncertainly-classified FRII), 8 (confidently-classified hybrid), 9 (uncertainly-classified hybrid)

FR-DEEP: 0 (FRI), 1 (FRII)

AT17: 0 (FRI), 1 (FRII), 2 (bent)

MiraBest Hybrid: 0 (confidently-classified hybrid), 1 (uncertainly-classified hybrid)

For examples of how to use CRUMB, please see its Github.

CRUMB 2.0 fixes an issue where images' class labels were not correctly assigned, removes several duplicate sources, and realigns several images to reduce ambiguity about which source is of interest. Use of CRUMB 1.0 is not recommended as a result of this labelling issue.