Datasets and protocols

Eight open wildlife re-identification datasets. Six come from WildlifeReID-10k with its official splits; CzechLynx uses its time-closed split; SalamanderID2025 (AnimalCLEF 2025) uses a split defined for this work because its query labels are not public.

The database split forms the retrieval gallery and is the only data used for fine-tuning. Query images are evaluated against it and never seen in training. Images are noisy and are not filtered.

Statistics

Dataset	Source	DB images	Q images	DB IDs	Q IDs	Med.	Max	Top 10 %	Single
Nyala	WildlifeReID-10k (NyalaData)	1,514	428	237	236	5	25	30 %	24 %
Hyena	WildlifeReID-10k (HyenaID2022)	2,499	630	256	256	6	72	35 %	2 %
Leopard	WildlifeReID-10k (LeopardID2022)	5,373	1,433	430	380	3	329	70 %	31 %
Sea star	WildlifeReID-10k (SeaStarReID2023)	1,759	428	95	95	14	30	15 %	0 %
Whale shark	WildlifeReID-10k (WhaleSharkID)	6,108	1,585	543	512	6	123	42 %	14 %
Turtle	WildlifeReID-10k (ZindiTurtleRecall)	9,721	3,082	2,265	2,251	2	92	39 %	3 %
Salamander	SalamanderID2025	1,138	246	584	221	1	11	31 %	64 %
CzechLynx	CzechLynx, time-closed split	27,836	11,924	319	319	26	2,465	61 %	0 %

Med. and Max are the median and maximum number of database images per individual. Top 10 % is the share of all database images held by the 10 % most photographed individuals. Single is the share of individuals with one image.

Note

Statistics follow the manuscript's dataset table; the page's copy is checked against the paper source by the repository's test suite.

Salamander split

For each individual photographed on at least two capture dates, all images of its latest date are queries and the remaining images form the database. Backgrounds were removed with text-prompted SAM 3 (prompt "Salamander", all detected instances merged because an occluding finger can split one animal into several instances).

Candidate budget

The default candidate budget is \(k = 250\); results are swept over \(k \in \{10, 50, 100, 250, 500, 1000\}\). Increasing \(k\) raises the chance that the correct individual is in the candidate list but costs proportionally more matching time.

Unseen-identity protocol (CzechLynx)

The matcher is trained on the training part of the CzechLynx time-open split (275 individuals). Evaluation uses the 44 individuals that appear in the test part of that split but not in its training part. For each of them, the earliest encounter forms the gallery (160 images) and later encounters the queries (2,081 images), which rules out encounter-level leakage. Budgets are \(k \in \{10, 50, 100, 160\}\), where 160 scores the whole gallery.

The protocol is generated by wildmatch prepare unseen-split, which groups images by encounter and assigns the earliest group to the gallery. It is never a random image-level split.

Known data problems

Examples of challenging images in the benchmark datasets. Raw frames that are overexposed (a), show too little of the animal to see its markings (b), contain no visible animal (c), are corrupted (d), or are blurred (e). Such images remain in the database and query sets of all methods.

The Data challenges page shows many more confirmed examples per problem, separates dataset noise from inherent difficulties such as night frames and occlusion, and reports how many images each audit flag caught and whether flagged queries are less accurate. Problems visible only in the masks, not in the raw photos, are described there in text.

Licensing

The photographs on this page are shown to illustrate the research, with attribution to their datasets. No dataset is redistributed: the code downloads or rebuilds each one from its published source.

Dataset	Source	License
CzechLynx photographs	CzechLynx (Kaggle)	CC BY 4.0
CzechLynx synthetic renders (demos)	CzechLynx, Zenodo record 17592004	CC BY 4.0
Hyena, Leopard, Sea star, Whale shark	WildlifeReID-10k (Kaggle), from LILA BC	CDLA-Permissive-1.0
Turtle	WildlifeReID-10k (Kaggle), from Zindi Turtle Recall	CC BY-SA 4.0
Nyala	WildlifeReID-10k (Kaggle)	no license stated by the source
Salamander	AnimalCLEF 2025 (Kaggle competition)	competition rules; shown with the dataset team's permission

WildlifeReID-10k additionally forbids commercial use and re-uploading, and asks users to attribute WildlifeReID-10k and each of its source datasets.