Pusa: Taxonomy and Conservation Insights

• Pusa is the accepted seal genus encompassing Saimaa and Baltic ringed seals, defined by unique pelage patterns used in non-invasive photo-identification and demographic studies.
• Advanced methods like SealID and NORPPA benchmark robust image-based re-identification, achieving Top‑1 accuracies up to 77.64% under challenging capture conditions.
• Integrated Bayesian models and computer vision pipelines inform conservation by linking pelage features with reproductive, behavioral, and mortality trends in ringed seal populations.

Searching arXiv for relevant papers on “Pusa” to ground the article in current literature. Pusa is the accepted genus name used in the cited literature for several ringed and closely related seals. In the sources considered here, the term appears primarily in zoological usage through studies of Pusa hispida saimensis, the Saimaa ringed seal endemic to Lake Saimaa, Finland, and Pusa hispida botnica, the Baltic ringed seal of the Bothnian Bay. These works treat Pusa as a taxonomic category, a conservation concern, and a technically demanding object of image-based individual re-identification and demographic modeling; the same string also appears in unrelated recent literature as the name of a video diffusion paradigm and as shorthand for the chickpea variety Pusa-372 (Nepovinnykh et al., 2022, Ersalman et al., 2024, Liu et al., 22 Jul 2025, Azimi et al., 2021).

1. Taxonomic scope and nomenclature

The zoological literature in the supplied corpus uses Pusa as the accepted genus. One study notes that the Saimaa ringed seal is referred to as Pusa hispida saimensis and that, although it may appear elsewhere as Phoca hispida saimensis, Pusa is now the accepted genus (Nepovinnykh et al., 2022). The same genus designation is used for the Baltic ringed seal, Pusa hispida botnica, in population-level work from the Bothnian Bay (Ersalman et al., 2024).

Within this frame, the corpus emphasizes two subspecific contexts. The Saimaa ringed seal is a freshwater subspecies of the ringed seal (Pusa hispida) and is only found in the Lake Saimaa, Finland. It is described as “one of the few existing freshwater seal species,” which makes it biologically distinctive. The Baltic ringed seal is presented as a Baltic endemic, with four recognized sub-populations—Bothnian Bay, Archipelago Sea, Gulf of Finland, and Gulf of Riga—and the Bothnian Bay sub-population is the largest, holding over 75% of all Baltic ringed seals (Nepovinnykh et al., 2022, Ersalman et al., 2024).

This nomenclatural point matters because much of the recent technical literature is organized around subspecies-level monitoring. In that literature, Pusa is not merely a classificatory label; it is the taxonomic anchor for conservation datasets, pelage-pattern models, and sex- and age-structured demographic analyses.

2. Pusa hispida saimensis as a conservation and monitoring system

The Saimaa ringed seal is described as an endangered species with around 400 individuals alive at the moment. Its entire global range is a fragmented lake system in eastern Finland, and effective conservation depends heavily on good demographic monitoring (Nepovinnykh et al., 2022). The supplied material repeatedly presents this subspecies as a canonical case in which non-invasive monitoring is essential.

A central biological property is the pelage. Ringed seals have a dark pelage ornamented by light grey rings, known as the fur patterns, and these are permanent and unique to each individual. Because the pelage pattern of the Saimaa ringed seal covers the whole surface area of a seal, photo-identification can in principle use any sufficiently visible body region. At the same time, it is impossible to see the full pattern from one image, so identification depends on matching partial observations across views (Nepovinnykh et al., 2022).

The same material also defines the technical difficulty of the problem. Large variation in poses is further exacerbated by the deformable nature of seals; the identifiable region may be limited; the contrast between the ring pattern and the rest of the pelage is low; wet and dry fur differ substantially; and image quality varies. These properties make Pusa hispida saimensis a difficult benchmark for automatic re-identification, but they also explain why the species is so prominent in computer-vision work on wildlife monitoring (Nepovinnykh et al., 2022).

Photo-identification is framed as a non-invasive alternative to tagging. Traditional tagging requires physical contact with the animal, which causes stress and may change the behavior of the animal. By contrast, Photo-ID provides tools to study migration, survival, dispersal, site fidelity, reproduction, health, population size, or density. For a small endemic population, the connection between individual recognition and mark–recapture analysis is direct (Nepovinnykh et al., 2022).

3. SealID and NORPPA: benchmark construction and pelage-based retrieval

The main public benchmark for Pusa hispida saimensis re-identification in the supplied literature is the SealID dataset. It was released for research purposes with 57 identified individuals and 2080 images in the main re-identification dataset, accompanied by segmentation masks and a separate pelage pattern patch dataset (Nepovinnykh et al., 2022).

SealID component	Size	Notes
Identified individuals	57	Saimaa ringed seals
Main re-ID images	2080	430 database, 1650 query
Pelage patch dataset	4599 patches	each 60×60 pixels

The dataset design is operationally motivated. The database set contains minimal sets of high-quality images per individual to cover visible pelage, while the query set contains all other images of the same individuals. Each query image is guaranteed to contain some part of a pattern that could be matched to the visible patterns in the database. The paper also defines a closed-set evaluation protocol using top-$k$ accuracy for $k \in \{1,3,5\}$, reflecting a semi-automatic workflow in which a biologist reviews a shortlist rather than relying on a single automatic decision (Nepovinnykh et al., 2022).

NORPPA—“NOvel Ringed seal re-identification by Pelage Pattern Aggregation”—casts identification as content-based image retrieval. Its pipeline consists of image preprocessing, seal instance segmentation, pelage pattern extraction, local feature extraction, Fisher Vector aggregation, and final ranking by cosine distance. In the described implementation, segmentation is done by Mask R‑CNN, pelage pattern extraction by a U‑net encoder–decoder method, local regions by HesAffNet, descriptors by HardNet, and global aggregation by Fisher Vectors (Nepovinnykh et al., 2022).

The reported results establish the method as the strongest baseline in the supplied corpus. On preprocessed images, HotSpotter reached Top‑1 69.39%, Top‑3 72.00%, and Top‑5 73.15%, whereas NORPPA reached Top‑1 77.64%, Top‑3 82.97%, and Top‑5 85.27%. In the NORPPA paper’s direct comparison, HotSpotter achieved Top‑1 61.87% and Top‑5 64.42%, while NORPPA achieved Top‑1 77.64% and Top‑5 85.27% (Nepovinnykh et al., 2022, Nepovinnykh et al., 2022).

These numbers clarify a common misunderstanding about wildlife re-identification benchmarks. The objective is not only fully automatic closed-set classification; it is also high-quality ranking under partial visibility, severe deformation, and strong background bias. The emphasis on Top‑5 and Top‑20 performance follows directly from how experts actually inspect candidate matches in monitoring programs (Nepovinnykh et al., 2022).

4. Geometry-aware unwrapping of Pusa pelage

A later development addresses a specific failure mode of pelage-based retrieval: pose-induced geometric distortion. “Unsupervised Pelage Pattern Unwrapping for Animal Re-identification” proposes a geometry-aware texture mapping approach that unwarps pelage patterns into a canonical UV space. For Saimaa ringed seals, the method uses SEEM for instance segmentation, Metric3Dv2 for surface normal estimation on the original unsegmented image, and an MLP with Fourier positional encoding and six dense layers with ReLU activations to learn a mapping

$g : (x, y) \mapsto (u, v).$

The resulting texture is produced with Delaunay triangulation and barycentric interpolation, and the entire UV-mapping stage is trained without ground-truth UV annotations in a self-supervised manner (Algasov et al., 18 Jun 2025).

The seal use case is explicit. Saimaa ringed seals have ring-shaped patterns on their fur, but the body is highly deformable, and grazing-angle views, stretching, and folding can make corresponding pelage regions appear geometrically inconsistent. The UV-space formulation is intended to preserve isometry between the 3D surface and the 2D texture space while making patterns more directly comparable across images (Algasov et al., 18 Jun 2025).

The evaluation reported for the seal subset is uniformly positive. On a subset described as 480 annotations and 146 individual animals, DoGHardNet + LightGlue improved from Top‑1 45.4% on original segmented images to 48.8% on unwrapped textures; SuperPoint + LightGlue improved from 43.3% to 48.0%; DISK + LightGlue improved from 27.9% to 33.3%; and ALFRE-ID improved from 45.0% to 47.6%. The paper summarizes the effect as up to a 5.4% improvement in re-identification accuracy (Algasov et al., 18 Jun 2025).

The method also states its own limitations. It is highly dependent on the quality of normal estimation, may degrade under low-light conditions or heavily obstructed views, and the current UV textures retain rotation and translation uncertainty in the UV plane. This indicates that pose normalization improves matching, but does not by itself remove every ambiguity in subtle Pusa pelage patterns (Algasov et al., 18 Jun 2025).

5. Pusa hispida botnica in Bayesian integrated population modeling

Where the Saimaa literature emphasizes individual recognition, the Baltic literature emphasizes population dynamics under changing environment and harvest. The integrated population model for Pusa hispida botnica is a Bayesian state-space model fitted to census and various demographic, reproductive, and harvest data from 1988 to 2023. It is explicitly sex- and age-structured, with age 0 pups, ages 1–4 sub-adults, and age 5+ adults, and uses an annual post-breeding census (Ersalman et al., 2024).

The model combines several observation processes. Annual late-April aerial surveys are treated as a negative-binomial observation of the hauled-out fraction of the population; hunting bag sizes and age–sex composition inform hunting mortality; bycatch samples inform non-hunting mortality structure; reproductive data from fetuses, placental scars, and corpus albicans inform time-varying fecundity. A haul-out submodel links the probability of being on ice to ice extent, and Finnish hunting mortality is also made ice-dependent (Ersalman et al., 2024).

The main quantitative result is that 20,000 to 36,000 ringed seals inhabited the Bothnian Bay in 2024, increasing at a rate of 3% to 6% per year. Reproductive rates have increased since 1988, leading to a substantial increase in the growth rate up until 2015. The re-introduction of hunting has since reduced the growth rate, and even minor quota increases are likely to reduce it further. The study also gives a critical harvest level of about 1,200 seals per year as the level at which growth would be zero under current conditions (Ersalman et al., 2024).

A particularly important clarification concerns monitoring interpretation. The results support the hypothesis that a greater proportion of the population hauls out under lower ice cover circumstances, leading to higher aerial survey results in such years. This means that aerial counts are not direct abundance measures with fixed detectability across years. In the supplied literature, Pusa hispida botnica therefore becomes a case study in why demographic inference must integrate behavior, environment, and observation process rather than rely on raw count trends alone (Ersalman et al., 2024).

6. Other scholarly uses of “Pusa”

Outside zoology, the supplied arXiv record uses “Pusa” in at least two unrelated ways. One is “PUSA V1.0,” a video diffusion paradigm based on vectorized timestep adaptation. In that work, Pusa is defined by a vectorized timestep variable $\bm{\tau}$, frame-aware flow matching, and LoRA fine-tuning on top of Wan2.1‑T2V‑14B. The paper states that it surpasses Wan‑I2V‑14B with $\leq 1/200$ of the training cost (\$500 vs. \(\geq \$100{,}000)) and $\leq 1/2500$ of the dataset size (4K vs. $\geq 10$M samples), reaches a VBench‑I2V total score of 87.32% versus 86.86% for Wan‑I2V‑14B, and retains text-to-video generation while adding zero-shot start-end frames and video extension capabilities (Liu et al., 22 Jul 2025).

The second is “Pusa” as shorthand for Pusa-372 in plant phenotyping. In the chickpea water-stress study, Pusa-372 is one of the two varieties examined and is explicitly characterized as stress-tolerant. For this variety, the dataset contains 3840 RGB images and 120 temporal sequences, and the CNN‑LSTM pipeline reaches ceiling-level performance around 97.5–97.8% accuracy, compared with a best time-invariant baseline of 84% (Azimi et al., 2021).

These usages are lexically identical but conceptually independent. In the supplied literature, Pusa therefore denotes, depending on context, a seal genus, a video-generation method, or a chickpea variety shorthand. The zoological usage remains the most cohesive across the corpus because it links taxonomy, pelage-based photo-identification, and demographic inference within a single research program spanning conservation biology, ecology, and computer vision (Nepovinnykh et al., 2022, Nepovinnykh et al., 2022, Algasov et al., 18 Jun 2025, Ersalman et al., 2024).