We developed a free pipeline for reliable individual identification in turtle populations by combining automated computer vision with targeted human validation. The pipeline processes raw images to automatically detect and extract biologically relevant regions such as the head, flippers, and carapace. These regions are compared across images to generate candidate matches, enabling the system to group observations from the same individual. An interactive application enables experts to efficiently verify or correct proposed matches, ensuring high accuracy while keeping manual effort focused and minimal.
The system supports two complementary modes of use:
It can be used to build new databases by automatically matching previously unassociated images and clustering them into individual turtles.
It can also be used to match new images against existing databases, enabling re-identification of known individuals as new data are collected over time.
Individual identification is crucial for establishing baseline population databases in new study sites, maintaining and expanding long-term monitoring programs. It supports conservation management by improving estimates of population size, survival, and movement, and by helping assess habitat use and the impact of protection measures. The system is also suitable for citizen science initiatives and for retrospective analysis of archival image collections, where new observations can be linked to existing individual records without prior marking.
The pipeline starts by detecting the turtle and segmenting its body parts (head, flippers, and carapace). For the flippers, we assign each a position (front or rear) and a side (left or right). The detector was trained on images of turtles both above and below water, and should work in both situations. We annotated the training images manually and trained a detector. Both the trained images and the detector are freely available online.
The photos below show one turtle observed in two different encounters. As with most of our databases and software, the images are freely available online (here and here). The top row shows the original images, while the bottom row shows the automatic segmentation.
turtle 1
turtle 2
Individual parts are matched by multiple algorithms based on deep networks. While it is relatively common to match turtle heads, our algorithms can also match both front and rear flippers. The extracted parts come from the images above. The lines between the images connect the same points on the turtle's body.
Our algorithm predicts which images are likely to depict the same turtle. Each prediction is assigned a similarity score, where a high similarity score means a higher probability of the images representing the same individual. These predictions are then fed into a clickable app, where experts can decide whether they believe the prediction is correct.
The verification can be performed by multiple experts. The algorithm compares them and makes note of any discrepancy. The correct predictions are then merged into the original database, preventing one turtle from having multiple identification codes (even if multiple tags have been applied).
If you would like to know more about the pipeline or to use it on your data, write me an email.