SeaTurtleID2022: A Long-Span Dataset for Reliable Sea Turtle Re-Identification

Lukáš Adam, Vojtěch Čermák, Kostas Papafitsoros, Lukas PicekWACV 2024

TL;DR

This paper introduces SeaTurtleID2022, the first large-scale, long-span dataset for sea turtle re-identification, collected over 13 years. It addresses the critical need for realistic dataset splits that account for temporal changes in appearance, demonstrating that random splits can lead to significant performance overestimation. The authors also propose and evaluate an end-to-end system for sea turtle re-identification, achieving 86.8% accuracy.

What Problem Does This Paper Address?

Identifying individual animals in the wild is crucial for wildlife research, enabling population monitoring, movement studies, and conservation efforts. Traditional methods often rely on manual identification, which is labor-intensive and scales poorly with the growing volume of photographic data. Automatic re-identification (re-ID) systems aim to automate this process. However, existing re-ID datasets often suffer from limitations: they may be small, lack temporal diversity, or use random splits that don't reflect real-world scenarios. This can lead to inflated performance metrics that don't translate to practical applications. Specifically, sea turtle re-identification faces challenges due to variations in appearance over time (e.g., growth, scratches, color changes) and the need for datasets that capture these long-term changes.

Methods

The paper introduces two novel, ecologically motivated dataset splitting strategies:

  • Time-aware closed-set split: This split ensures that reference (training) and query (test) images originate from different encounters or time periods, but all identities in the test set are known from the training set.
  • Time-aware open-set split: This is a more realistic scenario where the test set includes individuals not seen during training, reflecting the introduction of new animals to a population.

For re-identification, the authors explore:

  • Local feature-based methods: Using SIFT and Superpoint descriptors for matching.
  • Metric learning: Employing deep learning models like ArcFace and Triplet loss with a Swin-B backbone to learn discriminative embeddings.
  • Body-part segmentation: Investigating the importance of different body parts (head, flippers, full body) for re-identification, using instance segmentation models like Mask R-CNN, HTC, and Mask2Former.
  • End-to-end system: A pipeline that combines head detection (using Hybrid Task Cascade), feature extraction (with Swin-B ArcFace), and k-NN matching, with identity aggregation via majority voting within encounters.

Datasets

  • SeaTurtleID2022 (New Dataset): This is the primary contribution. It comprises 8729 high-resolution photographs of 438 unique loggerhead sea turtles, collected over 13 years (2010-2022) in Laganas Bay, Greece. It includes annotations for identity, encounter timestamps, and body part segmentation masks.
  • Existing Datasets: The paper also references and compares performance on other publicly available animal re-ID datasets (e.g., Cows2021, GiraffeZebraID, MacaqueFaces, BelugaID, WhaleSharkID, ZindiTurtleRecall) to highlight the novelty of SeaTurtleID2022, particularly its long-span and timestamped nature.

Key Insights and Ideas

  • The critical importance of time-aware splits: Random splits can drastically overestimate performance by including images from the same encounter or very similar conditions in both training and testing. This leads to models learning spurious correlations (like background similarities) rather than robust identity features.
  • Head is king for sea turtle re-ID: The study reveals that the turtle's head, with its unique scale patterns, is the most influential body part for accurate re-identification, significantly boosting performance compared to using the full body.
  • Long-span data is essential for realistic evaluation: The 13-year span of SeaTurtleID2022 allows for the study of appearance changes over time, a factor often overlooked in shorter-term datasets.
  • Encounter-based prediction improves accuracy: Aggregating predictions from multiple images of the same individual within an encounter (using majority voting) significantly enhances re-ID performance.

Key Results

  • Performance Inflation: Random splits showed accuracy inflation of up to 42.2% for full images, 53.8% for flippers, 45.8% for body, and 18% for head crops compared to time-aware splits.
  • Body Part Importance: Focusing solely on the head increased performance by 34.5% compared to using the full body. Flippers were found to be the least influential body part.
  • End-to-End System Performance: The proposed end-to-end system achieved an accuracy of 86.8% on SeaTurtleID2022 using time-aware splits. This is a substantial improvement over a naive approach using full images without body part information or encounter knowledge (17.1% accuracy).
  • Segmentation Performance: The Mask2Former model with a Swin-B backbone achieved a competitive COCO mAP of 0.896 for instance segmentation.

Limitations and Future Work

  • Limitations:
    • The dataset is collected from a single location (Laganas Bay), which might limit generalizability to other populations or environments.
  • While the dataset is large-scale, further expansion with more individuals and longer time spans would be beneficial.
  • The current end-to-end system relies on head detection; failure to detect a head means no identity prediction.
  • Future Work:
    • The authors suggest that SeaTurtleID2022 can be used to evaluate other computer vision tasks like object detection and semantic segmentation.
  • The proposed end-to-end system is designed to be transferable to other species.
  • Further research could explore more sophisticated methods for handling appearance changes over time and integrating information from multiple body parts.

Summary updated

BibTeX
@inproceedings{adam2024seaturtleid,
  author = {Lukáš Adam and Vojtěch Čermák and Kostas Papafitsoros and Lukas Picek},
  title = {{SeaTurtleID2022: A Long-Span Dataset for Reliable Sea Turtle Re-Identification}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2024},
}
Figure