imageomics/cfr-wildlife-masks

CFR Wildlife Masks

This repository accompanies CounterFactual Retrieval for Evidence-Aware Individual Animal Identification and gives the community access to 330,157 foreground masks across 33 wildlife re-identification datasets, generated from 351,595 annotations. These masks support foreground-based retrieval and analysis of the visual evidence used to identify individual animals. Coverage is planned to expand to most datasets supported by WildlifeDatasets.

The release contains generated masks and the metadata needed to match them to source images. For supported datasets, original images and annotations can be downloaded using WildlifeDatasets; source links are listed below. Datasets that already supply masks or foreground cutouts are excluded.

Datasets and sources

Annotations counts all processed inputs; masks counts accepted outputs, with at most one mask per annotation. Source and paper links credit the original data providers and describe their collections.

Dataset	Source / associated paper	Annotations	Accepted masks
AAUZebraFish	Dataset · Haurum et al. (2020)	6,672	6,672
ATRW	Dataset · Li et al. (2019)	5,415	5,393
AerialCattle2017	Dataset · Andrew et al. (2017)	46,340	30,016
AmvrakikosTurtles	Dataset · Adam et al. (2025)	200	198
BelugaID	Dataset	5,902	4,914
BrownBearHeads	Dataset · Rosenberg et al. (2026)	72,939	72,911
CTai	Dataset · Freytag et al. (2016)	5,078	5,039
CZoo	Dataset · Freytag et al. (2016)	2,109	2,106
CattleMuzzle	Dataset · Mirror used · Shojaeipour et al. (2021)	2,899	2,891
Cows2021	Dataset · Gao et al. (2021)	8,670	7,912
DogFaceNet	Dataset · Mougeot et al. (2019)	8,363	8,341
FriesianCattle2017	Dataset · Andrew et al. (2017)	940	938
GiraffeZebraID	Dataset · Parham et al. (2017)	6,925	6,911
Giraffes	Miele et al. (2021); data availability	1,393	1,392
HappyWhale	Dataset · Patton et al. (2023)	51,033	50,930
HumpbackWhaleID	Dataset · Cheeseman et al. (2022)	33,321	32,099
HyenaID2022	Dataset	3,129	3,027
IPanda50	Dataset · Wang et al. (2021)	6,874	6,874
Leopard †	Related collection	4,564	4,548
LionData	Dataset · Dlamini and van Zyl (2020)	740	694
MPDD	Dataset · He et al. (2023)	1,657	1,635
MacaqueFaces	Dataset · Witham (2018)	6,280	6,280
Melops	Dataset · Sørdalen et al. (2026)	24,578	24,578
NOAARightWhale	Dataset · Khan et al. (2022)	11,469	11,424
NyalaData	Dataset · Dlamini and van Zyl (2020)	1,942	1,942
OpenCows2020	Dataset · Andrew et al. (2021)	4,736	3,252
ReunionTurtles	Dataset · Adam et al. (2025)	336	336
SaimaaSeal ‡	Nonpublic source collection	2,934	2,857
SeaStarReID2023	Dataset · Wahltinez and Wahltinez (2024)	2,187	2,159
StripeSpotter	Dataset · Lahiri et al. (2011)	824	824
WhaleShark	Dataset · Holmberg et al. (2009)	7,693	7,654
ZakynthosTurtles	Dataset · Adam et al. (2025)	160	160
ZindiTurtleRecall	Dataset	13,293	13,250
Total		351,595	330,157

Source-provided masks or foregrounds (15 datasets): For BalearicLizard, BirdIndividualID, CHIRP, Chicks4FreeID, CoBRAReIdentificationYoungstock, CzechLynx, FriesianCattle2015, GCN_ID, NDD20, PrimFace, RedBeeReID, RotwildID_Faces, SMALST, SeaTurtleID2022, and WildRaptorID, use the masks, segmentation annotations, or foreground images supplied by the original datasets. Together with the 33 datasets in this release, this gives 48 datasets with generated or source-provided masks/foregrounds. Source formats and annotation coverage vary by dataset.

† The processed Leopard collection overlaps LeopardID2022 but is not an exact copy; the linked collection is related provenance.

‡ SaimaaSeal is the 2,934-annotation norppa/Wildbook collection used by CFR, not the separate public SealID dataset.

Files and use

Download any dataset independently from data/<dataset>/. Every folder contains the same three files:

File	Contents
masks.json.gz	Compressed JSON containing an annotations list. Accepted records contain a binary mask encoded as COCO RLE, its area and bounding box. Rejected records have present=false and segmentation=null.
metadata.parquet	One row per processed annotation: source-image reference, crop geometry, acceptance status, prompt and selected-mask confidence. Join to masks using the string annotation_id.
summary.json	Annotation/mask counts, rejection reasons, prompts, selection rule, and model/pipeline versions.

Rejected records are retained for coverage accounting. The Hub's annotations split displays the metadata inventory, not a new evaluation split. Use the source datasets' identity labels and split protocols. HappyWhale contains its 51,033 originally labeled images; NOAARightWhale includes its train/test inventory.

Coordinates and matching

Masks use the SAM input grid, which may be cropped or resized relative to the source photograph. RLE size is [height, width]; metadata dimensions use [width, height].

Fields	Meaning
annotation_id, image_id	Extraction identifiers; original source identifiers are retained in source_* fields when available.
input_image, input_file_size_wh, input_crop_xyxy, input_size_wh	Loaded image, its dimensions, the half-open crop [x0,y0,x1,y1], and resulting mask dimensions.
source_image, source_size_wh, source_crop_xywh, source_crop_xyxy	Source path and geometry when known; xywh means [x,y,width,height]. Missing values are null.
source_transform, source_transform_status	Image transformation before mask extraction. exact records a known source transformation; recorded retains prior preprocessing without claiming that its historical rounding has been reproduced.
accepted, status, confidence, mask_area	Mask availability, outcome/rejection reason, selected SAM score, and foreground pixel count.
prompts, selected_prompt, selection_policy, fallback_applied	Prompt and candidate-selection provenance.
input_rgb_sha256	Fingerprint of the decoded RGB inference pixels, not the image file.

Paths are relative and retain their archive-folder prefixes. prepared_518 refers to the resized input cache used during extraction; use source_image to find upstream photographs. For Leopard and WhaleShark, source paths start inside the COCO images/ directory. norppa_database_crops identifies the supplied Saimaa PNG crops. StripeSpotter masks use the full image, not the annotated flank box. Align these coordinate frames before overlaying masks or computing IoU.

For ZakynthosTurtles, apply the recorded EXIF orientation and crop the source photograph using source_crop_xyxy. One SeaStarReID2023 image also requires EXIF rotation. When source_transform starts with exif_transpose_, source dimensions and crop coordinates refer to the image after that rotation. DogFaceNet, MPDD, and CattleMuzzle retain their source dimensions.

Load a mask

Install huggingface_hub, pandas, pyarrow and pycocotools.

import gzip
import json
import pandas as pd
from huggingface_hub import hf_hub_download
from pycocotools import mask as mask_utils

repo = "ankitkupadhyay/cfr-wildlife-masks"
dataset = "NyalaData"
mask_path = hf_hub_download(repo, f"data/{dataset}/masks.json.gz", repo_type="dataset")
meta_path = hf_hub_download(repo, f"data/{dataset}/metadata.parquet", repo_type="dataset")
with gzip.open(mask_path, "rt") as stream:
    records = json.load(stream)["annotations"]
metadata = pd.read_parquet(meta_path).set_index("annotation_id")
record = next(record for record in records if record["present"])
info = metadata.loc[str(record["annotation_id"])]
rle = dict(record["segmentation"])
rle["counts"] = rle["counts"].encode("ascii")
binary_mask = mask_utils.decode(rle)  # uint8 [height, width], values 0/1
assert binary_mask.shape == tuple(reversed(info["input_size_wh"]))

The root SHA256SUMS file verifies downloaded file integrity: run sha256sum -c SHA256SUMS after downloading the complete repository.

Mask generation

Meta SAM3 is prompted for the appropriate species or body part on the recorded annotation crop or supplied image. Among candidates whose bounding boxes contain the input center, select the mask with the most foreground pixels. Break ties by confidence, then candidate order. The center pixel does not need to lie inside the foreground mask.

For HappyWhale and NOAARightWhale, if no candidate box contains the center, select the highest-confidence valid mask; break ties by foreground pixel count, then candidate order. No valid candidate means rejection. GiraffeZebraID and HappyWhale route prompts by annotation species; where multiple prompts are used, their candidates share the same selection rule.

The recorded SAM thresholds are confidence > 0.3 and mask probability > 0.5. No additional coverage threshold, hole filling or speck removal is applied. Each summary records the exact model revision and pipeline version. This collection uses the revised extraction procedure and is distinct from the frozen masks used for the paper's reported experiments.

Limitations

These are automatic predictions, not human ground truth. Accepted masks can contain the wrong animal or imperfect boundaries, particularly with multiple animals or occlusion. Acceptance counts measure extraction coverage, not segmentation accuracy. Validate masks and coordinate alignment for the intended application.

Authors and affiliations

Ankit K. Upadhyay¹, Aleksandr Algasov²˒³, Ekaterina Nepovinnykh², Fedor Zolotarev⁴, Tuomas Eerola², Heikki Kälviäinen²˒³, Charles V. Stewart¹, Lukáš Picek⁵˒⁶.

¹ Rensselaer Polytechnic Institute · ² LUT University · ³ Brno University of Technology · ⁴ University of Helsinki · ⁵ Massachusetts Institute of Technology · ⁶ University of West Bohemia.

Contact: Ankit K. Upadhyay.

Licensing information

Masks provided in this repository are licensed under CC BY 4.0. Source datasets and third-party metadata retain their respective licenses and terms.

Citation

Please cite CFR and the relevant source datasets/papers linked above, WildlifeDatasets, and SAM3.

Provisional CFR citation supplied by the authors:

@inproceedings{upadhyay2027cfr,
  author = {Ankit Kumar Upadhyay and Aleksandr Algasov and Ekaterina Nepovinnykh
            and Fedor Zolotarev and Tuomas Eerola and Heikki Kälviäinen
            and Charles V. Stewart and Lukáš Picek},
  title = {CounterFactual Retrieval for Evidence-Aware Individual Animal Identification},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  year = {2027},
  note = {Accepted}
}

Acknowledgements

Supported by the Imageomics Institute, funded by the US National Science Foundation's Harnessing the Data Revolution program under Award #2118240. Opinions, findings, conclusions and recommendations are those of the authors and do not necessarily reflect the views of the National Science Foundation.