HotSpotter: Fast Animal Re-Identification

• HotSpotter is a species-agnostic system that identifies animals by extracting and matching distinctive keypoints (hotspots) from user-specified image regions.
• It employs affine-covariant detection and 128-dimensional RootSIFT descriptors with geometric verification (e.g., RANSAC) to robustly match partial patterns.
• The algorithm supports both one-vs-one and one-vs-many matching modes, offering high accuracy and efficiency in wildlife re-identification studies.

HotSpotter is a fast, accurate algorithm for identifying individual animals against a labeled database. It is not species specific, has been applied to Grevy’s and plains zebras, giraffes, leopards, and lionfish, and is built around the extraction and matching of distinctive keypoints, or “hotspots,” from animal imagery (Crall et al., 25 Aug 2025). In the animal re-identification literature, it functions as a general-purpose, local-feature-based instance recognition system: a query image is compared to a database of labeled images, scored through keypoint correspondences and geometric verification, and returned as a ranked list of candidate identities (Nepovinnykh et al., 2022).

1. Scope and problem setting

HotSpotter belongs to the class of patterned-species re-identification systems. Its core use case is identification of individuals from naturally occurring markings such as stripes, spots, rosettes, blotches, or other visually distinctive coat or skin patterns. The method is explicitly species-agnostic at the descriptor and matching level, rather than being tuned to a single morphology or marking type (Crall et al., 25 Aug 2025).

The standard operational setting is database retrieval rather than closed-set classification. A labeled reference database is maintained, and each new query image is ranked against that database. This makes HotSpotter suitable for wildlife monitoring contexts in which the number of images per individual may be limited, and where adding new identities does not require retraining a parametric model. The SealID study characterizes it as a generic animal re-identification algorithm previously used for species such as zebras, giraffes, and big cats, and emphasizes that it does not require large per-individual training data (Nepovinnykh et al., 2022).

A common misunderstanding is to equate “species-agnostic” with “fully automatic.” In the reported configurations, HotSpotter still depends on a manually specified rectangular region of interest around the animal, and in flank-based leopard identification this ROI corresponds to manual selection of the animal flank (Crall et al., 25 Aug 2025, Matachana, 2024). Its generality lies in the matching machinery, not in the elimination of annotation or cropping.

2. Image representation and hotspot extraction

The image representation is local, affine-covariant, and explicitly keyed to repeatable pattern elements. A user manually specifies a rectangular ROI around the animal; only this patch is analyzed. Within that ROI, HotSpotter detects keypoints using the Hessian-Hessian detector, which is described as faster and more stable than the Difference-of-Gaussians or Hessian-Laplace detectors used in other systems (Crall et al., 25 Aug 2025).

Each detected region is fitted with an ellipse, yielding an affine region. Orientation is assigned as vertical, with the gravity vector pointing down, rather than by a gradient-based rule, in order to increase discrimination. For each affine region, a 128-dimensional RootSIFT descriptor is extracted. The system stores locations, affine regions, and descriptors for later matching (Crall et al., 25 Aug 2025).

In the re-identification literature, the term “hot spots” denotes the most distinctive and repeatable keypoints on an animal pattern. The SealID baseline description characterizes HotSpotter as a SIFT-based algorithm using viewpoint invariant descriptors and a scoring mechanism that emphasizes these distinctive keypoints. The same description notes affine invariance for robustness to viewpoint changes and geometric verification, often via RANSAC, after descriptor matching (Nepovinnykh et al., 2022).

This architecture makes the representation fundamentally part-based. Matching can succeed on a subset of visible markings rather than on the full body surface. That design is especially relevant in wildlife imagery, where occlusion, pose variation, and partial visibility are routine. The leopard comparison explicitly reports that HotSpotter can succeed even if one rosette matches, and describes it as robust to occlusions (Matachana, 2024).

3. Matching architectures and scoring rules

HotSpotter implements two matching approaches: sequential image scoring and fast nearest-neighbor search (Crall et al., 25 Aug 2025).

In the sequential one-vs-one mode, a query image is compared independently with each database image. Descriptor matches are filtered using the classic SIFT ratio test,

$$r_{i,j} = \frac{\|d_i - q_{j_2}\|^2}{\|d_i - q_j\|^2},$$

where $d_i$ is a descriptor from the database image and $q_j, q_{j_2}$ are its two closest descriptors in the query. The retained correspondences are aggregated into an image similarity score,

$$Sim(\mathcal{M}_D) = \sum_{(i,j,r_{i,j}) \in \mathcal{M}_D} r_{i,j}.$$

The top matches are then spatially reranked through RANSAC-based filtering, and label scoring aggregates evidence across multiple images with the same individual identity (Crall et al., 25 Aug 2025).

In the fast one-vs-many mode, all database descriptors are indexed jointly in a single k-d tree forest. For each query descriptor, the system retrieves the $k+1$ nearest database descriptors and assigns competitive scores derived from Local Naive Bayes Nearest Neighbor. The reported scoring options include

$$\delta_{LNBNN}(q, d_p, d_{k+1}) = \|d_{k+1} - q\|^2 - \|d_p - q\|^2,$$

$$\delta_{ratio}(q, d_p, d_{k+1}) = \frac{\|d_{k+1} - q\|^2}{\|d_p - q\|^2},$$

and

$$\delta_{lnrat}(q, d_p, d_{k+1}) = \ln\left(\frac{\|d_{k+1} - q\|^2}{\|d_p - q\|^2}\right).$$

A count score, $\delta_{count}=1$, is also considered. As in one-vs-one matching, scores are attributed to database images and labels, then refined by spatial reranking and label scoring (Crall et al., 25 Aug 2025).

The conceptual distinction between the two modes is computational rather than representational. The one-vs-one pipeline preserves pairwise isolation at scoring time; the one-vs-many pipeline exploits global nearest-neighbor indexing and competitive distinctiveness scoring. The latter is reported to be as accurate as the former, but much faster (Crall et al., 25 Aug 2025).

4. Cross-species evaluations and computational profile

The broad empirical case for HotSpotter is built around heterogeneous patterned-species datasets. Reported evaluations include large and challenging sets such as 1047 Grevy’s zebra images, 824 plains zebra images, and smaller giraffe, jaguar, and lionfish sets (Crall et al., 25 Aug 2025).

Dataset or species	Reported result	Source
Grevy’s zebras	Up to 95% correct top-1 label matches; 98% in the top 5 suggestions	(Crall et al., 25 Aug 2025)
Plains zebras	Well over 99% correct matches	(Crall et al., 25 Aug 2025)
Giraffe, jaguar, lionfish	100% correct	(Crall et al., 25 Aug 2025)

These results are paired with a pronounced change in retrieval speed when moving from sequential to indexed matching. For Grevy’s zebras, one-vs-one matching required 38–78 seconds per query, whereas one-vs-many required 2–4 seconds per query; for plains zebras, one-vs-one required 38–41 seconds per query (Crall et al., 25 Aug 2025). The same source reports that product quantization further reduced memory usage and improved speed for massive databases, with little to no loss in accuracy for appropriate parameter choices.

The reported comparative baseline in these studies is Wild-ID. HotSpotter is described as outperforming Wild-ID and succeeding on jaguar and lionfish cases where Wild-ID failed (Crall et al., 25 Aug 2025). This suggests that the gain is not reducible to a single species or marking regime, but arises from the combination of affine-covariant region detection, RootSIFT description, competitive scoring, and geometric reranking.

5. Benchmark behavior on seals and leopards

Modern benchmark studies position HotSpotter as a strong baseline whose performance depends substantially on image preparation and species-specific visual structure. On the SealID dataset of Saimaa ringed seals, evaluation uses top-$k$ accuracy,

$d_i$0

with Top-1, Top-3, and Top-5 reported (Nepovinnykh et al., 2022).

On SealID, HotSpotter achieved 61.87% Top-1 on raw images and 69.39% Top-1 after preprocessing with segmentation and tone mapping. In the same study, NORPPA scored 49.52% Top-1 on raw images and 77.64% Top-1 after preprocessing (Nepovinnykh et al., 2022). The immediate implication is that preprocessing improves HotSpotter, but is even more important for NORPPA. More specifically, HotSpotter outperformed NORPPA without preprocessing, whereas NORPPA outperformed HotSpotter by a notable margin after preprocessing (Nepovinnykh et al., 2022).

On a leopard identification dataset, HotSpotter is compared directly with deep metric-learning systems under open-set evaluation. The reported figures are DT5AP $d_i$1 and T5RMD $d_i$2 for HotSpotter, versus DT5AP $d_i$3 and T5RMD $d_i$4 for the best modified CosFace model (Matachana, 2024). The same comparison reports that HotSpotter provides the highest accuracy in both metrics on that dataset, while still requiring manual ROI selection (Matachana, 2024).

These two case studies clarify the algorithm’s empirical profile. On strongly patterned species such as leopards, a mature local-feature pipeline can remain accuracy-leading relative to newer deep-learning baselines. On more subtle or low-contrast patterns such as Saimaa ringed seals, HotSpotter remains competitive but can be overtaken by a species-specific method once segmentation and normalization are optimized (Nepovinnykh et al., 2022, Matachana, 2024).

6. Limitations, failure modes, and relation to newer methods

The principal limitations reported for HotSpotter arise from image conditioning, annotation workflow, and the local nature of its representation. The 2025 HotSpotter study attributes most failures to incorrectly labeled database images, ROIs covering multiple animals, keypoints on background features rather than the animal, significant pose variation, and poor image quality in focus, contrast, or resolution (Crall et al., 25 Aug 2025).

The SealID analysis adds a second set of limitations tied to pattern visibility. Because the ringed seals’ pelage pattern covers the whole body, no single image can capture the full pattern, and performance can drop when query and database images show different body regions. The same study also emphasizes low contrast between ring patterns and the rest of the pelage, plus background clutter, different illuminations, and variable image quality (Nepovinnykh et al., 2022). This indicates that HotSpotter’s local descriptors are not uniformly equally informative across species and imaging regimes.

A further misconception is that deep learning has categorically displaced keypoint-based wildlife identification. The comparative record in the supplied literature is more specific. On the leopard dataset, the modified CosFace model is described as a substantial advancement in applying deep learning to patterned wildlife identification, yet it still does not surpass the performance of the SIFT-based HotSpotter algorithm (Matachana, 2024). At the same time, deep models offer automatic ROI detection, end-to-end learned embeddings, and a more scalable training-based route for open-set deployment, whereas HotSpotter retains manual ROI extraction and a handcrafted matching pipeline (Matachana, 2024).

HotSpotter therefore occupies a distinct position in the methodological landscape. It is a species-agnostic, local-feature, retrieval-based system whose strongest reported advantages are high performance on clearly patterned species, robustness to occlusion, and the absence of a need for large per-individual training sets (Nepovinnykh et al., 2022, Matachana, 2024). Its principal constraints are manual ROI specification, sensitivity to preprocessing and image presentation, and reduced advantage when patterns are subtle or when species-specific models exploit domain-adapted preprocessing (Nepovinnykh et al., 2022).