josauder/UCSD-mosaics-mirror

UCSD Mosaics (Mirror)

This is a HuggingFace mirror of the GT-Clean version of the UCSD Mosaics dense semantic segmentation dataset. It contains 512x512 patches sliced from large-area coral reef mosaics gathered across 16 dive sites in Palau and the Northern Line Islands (Palmyra, Fanning, Kingman, Jarvis), with single-channel uint8 segmentation masks over 34 named benthic classes.

Pixel value 0 is the unlabeled / unidentified ignore label (absorbing the source legend's Unidentified class, raw mosaic ID 34); it is not itself counted as one of the 34 classes.

This mirror is intended to make the dataset easier to consume with the datasets library; the source data and methodology are unchanged. Please cite the original works listed below.

Dataset Structure

The dataset is organized as one split per dive site (16 splits total). Splits are named after the lowercased site ID (for example, PALWave13 -> palwave13). Each row contains the patch image, the remapped segmentation mask, the original filename, the site (original case), and the patch's original_split (train or test) from the source GT-Clean release.

Quick start

from datasets import load_dataset, concatenate_datasets

repo = "josauder/UCSD-mosaics-mirror"

# Load a single site:
site_ds = load_dataset(repo, split="fr3")
sample = site_ds[0]
image = sample['image']             # PIL RGB image, 512x512
label = sample['label']             # PIL 'L' image, 512x512, values in 0..34 (0 = ignore)
filename, site = sample['filename'], sample['site']

# Load all sites as a DatasetDict (split name -> Dataset):
all_sites = load_dataset(repo)

# Concatenate every site into one big Dataset:
full = concatenate_datasets(list(all_sites.values()))

# Recreate the source GT-Clean train/test split using `original_split`:
train = full.filter(lambda ex: ex['original_split'] == 'train')
test  = full.filter(lambda ex: ex['original_split'] == 'test')

Sites

Site	Split name	Patches	of which train	of which test
FR3	fr3	220	184	36
FR5	fr5	262	225	37
FR7	fr7	415	353	62
FR9	fr9	310	258	52
PAL132	pal132	323	274	49
PAL239	pal239	292	244	48
PAL36	pal36	259	221	38
PAL69	pal69	304	269	35
PALStrawn	palstrawn	477	413	64
PALWave13	palwave13	314	272	42
PALWave14	palwave14	318	275	43
PALWave37	palwave37	297	251	46
PALWave38	palwave38	242	198	44
PALWave39	palwave39	254	217	37
PALWave4	palwave4	145	126	19
PALWave40	palwave40	238	194	44
Total		4670	3974	696

Classes

Label masks are single-channel uint8 PNGs with values in 0..34. Pixel value 0 is the unlabeled / unidentified ignore label and is not listed below. The 34 named classes use IDs 1..34 and are taken from the source legend.csv, skipping the catch-all Unidentified row (raw mosaic ID = 34), which is folded into the ignore label. The RGB colors below are the BGR palette from the source README converted to RGB.

Id	Name	Morphology	Color (RGB)
1	Acropora (branching)	Branching	#9F12A7 (159, 18, 167)
2	Acropora (corymbose)	Corymbose	#5C1BB4 (92, 27, 180)
3	Acropora (plating)	Tabular	#E98B68 (233, 139, 104)
4	Astreopora myriophthalma	Sub_massive	#87C631 (135, 198, 49)
5	Clavularia	Fleshy_invert	#1ACF62 (26, 207, 98)
6	Corallimorph	Corallimorph	#85D076 (133, 208, 118)
7	Dictyosphaeria	Encrusting_macroalgae	#5A769E (90, 118, 158)
8	Favia matthai	Massive	#A6480C (166, 72, 12)
9	Favia stelligera	Sub_massive	#EE4F45 (238, 79, 69)
10	Favites (encrusting)	Encrusting	#31C351 (49, 195, 81)
11	Favites (submassive)	Sub_massive	#34ECDD (52, 236, 221)
12	Fungia	Free_living	#DEC8A0 (222, 200, 160)
13	Halimeda	Erect_macroalgae	#D83FFF (216, 63, 255)
14	Halomitra pileus	Free_living	#075E10 (7, 94, 16)
15	Hydnophora exesa	Plating	#402FE2 (64, 47, 226)
16	Hydnophora microconos	Massive	#056CB7 (5, 108, 183)
17	Leptastrea	Encrusting	#C1FC37 (193, 252, 55)
18	Lobophyllia	Massive	#C49A93 (196, 154, 147)
19	Montastrea curta	Massive	#A54EE9 (165, 78, 233)
20	Montipora (encrusting)	Encrusting	#5F196C (95, 25, 108)
21	Montipora (plating)	Plating	#2EDDB8 (46, 221, 184)
22	Pasmmocora	Encrusting	#91CD36 (145, 205, 54)
23	Pavona (submassive)	Sub_massive	#D2650E (210, 101, 14)
24	Pavona varians	Encrusting	#E6E8C7 (230, 232, 199)
25	Platygyra	Massive	#670A42 (103, 10, 66)
26	Pocillopora	Corymbose	#3BE4A1 (59, 228, 161)
27	Pocillopora eydouxi	Corymbose	#68026C (104, 2, 108)
28	Porites (massive)	Massive	#7F310D (127, 49, 13)
29	Porites rus	Massive	#2663BA (38, 99, 186)
30	Porites superfusa	Encrusting	#F68C61 (246, 140, 97)
31	Soft coral	Soft	#CA722C (202, 114, 44)
32	Stylophora pistillata	Corymbose	#761F24 (118, 31, 36)
33	Turbinaria reniformis	Plating	#8F4D92 (143, 77, 146)
34	Zooanthid	Fleshy_invert	#0E64BC (14, 100, 188)

License & usage notice

The source dataset is distributed for research purposes only and should not be redistributed beyond research use. By using this mirror you agree to the same terms and to cite all of the works listed in the Citation section.

Citation

Please cite all of the following works when using this dataset:

Original large-area mosaics:

@article{edwards2017largearea,
  title   = {Large-area imaging reveals biologically driven non-random spatial patterns of corals at a remote reef},
  author  = {Edwards, Clinton B. and others},
  journal = {Coral Reefs},
  volume  = {36},
  year    = {2017}
}

Patch splitting and dense annotations:

@article{alonso2019coralseg,
  title   = {CoralSeg: Learning coral segmentation from sparse annotations},
  author  = {Alonso, I{\~n}igo and Yuval, Matan and Eyal, Gal and Treibitz, Tali and Murillo, Ana C.},
  journal = {Journal of Field Robotics},
  volume  = {36},
  year    = {2019}
}

GT-Clean variant (corrupted-mask removal) shipped here:

@inproceedings{raine2024human,
  title     = {Human-in-the-loop segmentation of multi-species coral imagery},
  author    = {Raine, Scarlett and Marchant, Ross and Kusy, Brano and Maire, Frederic and Sunderhauf, Niko and Fischer, Tobias},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages     = {2723--2732},
  year      = {2024}
}