Abstract
Automated wildlife reidentification has attracted increasing attention in recent years as it provides a non-invasive tool to identify and to track individual wild animals over time. In this paper, the first steps are taken towards the automatic photo-identification of the Ladoga ringed seals (Pusa hispida ladogensis). A method is proposed that takes a sequence of images, each containing multiple individuals as the input, and produces cropped images of seals grouped based on one certain individual per group. The method starts by detecting each seal from the images and proceeds to matching the individual seals between the images. It is shown that high grouping accuracy can be obtained with a general-purpose image retrieval method on an image sequence taken from the same location within a relatively short period of time. Each resulting group contains multiple images of one individual with slightly different variations, for example, in pose and illumination. Utilizing these images simultaneously provides more information for the individual re-identification compared to the traditional approach, i.e., which utilizes just one image at a time. It is further demonstrated that a convolutional neural network based method can be used to extract the unique pelage patterns of the seals despite the low contrast. Finally, a method is proposed and experiments with the novel Ladoga ringed seals data are carried out to provide a proof-of-concept for the individual re-identification.
Introduction
Wildlife monitoring often relies on "photo-identification"—recognizing individual animals by their unique natural markings. While this works well for high-contrast animals like zebras, the Ladoga ringed seal (Pusa hispida ladogensis) presents significant challenges for computer vision. Their "ring" patterns are faint and often nearly invisible depending on lighting and wetness. Furthermore, unlike their solitary Saimaa cousins, Ladoga seals huddle together in social groups. A single photo might contain a dozen overlapping individuals, making simple "one-seal-per-image" assumptions fail. Additionally, seals are "squishy," meaning their patterns deform as they move, making a single snapshot an unreliable source of truth. This research introduces an end-to-end framework to move from raw, crowded field photos to successful individual identification.
Methodology
The proposed pipeline consists of four distinct stages:
1. Instance Segmentation
Using Mask R-CNN with a ResNet-101-FPN backbone, the system detects and outlines every individual seal in a crowded image, separating them from the rocky background and each other.
2. Individual Grouping
Instead of treating every crop as a separate query, the system uses a ResNet-101 with Generalized Mean (GeM) pooling to group images of the same seal taken during the same session. Since seals remain relatively stationary between consecutive shots, visual similarity serves as a strong proxy for identity within a short timeframe.
3. Pattern Extraction
A U-Net architecture transforms the low-contrast fur crops into binary masks of the ring patterns. This model was pre-trained on the Saimaa ringed seal, demonstrating that features generalize across subspecies.
4. Re-identification
The extracted patterns are broken into patches. A CNN trained with SphereFace loss creates embeddings for these patches, which are then aggregated into Fisher Vectors. These vectors represent the "identity" of the seal group and are compared against a database using cosine similarity.
Results and Discussion
The study utilized a novel dataset of Ladoga ringed seals collected in 2019 and 2020. Key performance metrics include:
- Segmentation: The ResNet-101-FPN backbone achieved a 71.1% mAP.
- Grouping: The system achieved a 95.4% accuracy (Rand Index) in correctly clustering images of the same individual within a sequence.
- Re-ID Accuracy: The proposed pipeline (Grouping + Pattern Extraction) achieved a 93.62% Top-1 accuracy, a massive jump from the 53.18% achieved by standard image-to-image matching without grouping.
The most significant insight is that grouping matters more than the matching algorithm itself. By combining information from multiple angles and poses within a sequence, the system can "see" parts of a pattern that might be obscured in a single frame. While the U-Net only successfully extracted a "satisfactory" pattern from about 42% of images, the grouping mechanism mitigates this by requiring only one good extraction per group.
Conclusion
This research provides a robust proof-of-concept for the automated identification of Ladoga ringed seals. By leveraging short-term image sequences and transfer learning from related subspecies, the system overcomes the hurdles of low-contrast markings and social crowding. Future work will focus on improving pattern extraction rates and testing scalability against larger population databases.
Reference: Nepovinnykh, E., et al. (2022). Matching individual Ladoga ringed seals across short-term image sequences. Mammalian Biology.
Summary updated