Marine Conservation Computer Vision

Business Problem

Marine biologists tracking sea turtle populations rely heavily on manual photo-identification to distinguish individual animals and assess health and injury patterns, a slow, labour-intensive process that limits how much of a population researchers can realistically monitor.

Customer / User Context

The end users are conservation researchers who currently review and annotate field imagery by hand. Their constraint isn't a lack of imagery, it's the analyst time required to process it, which caps how much of a population can be monitored in a season.

Approach

Built an instance segmentation pipeline to automatically isolate a sea turtle's distinct anatomical regions, flippers, head/body and tail, from photographs, as a foundation for downstream identification and injury-assessment tasks that currently depend on manual annotation. Implemented and compared three segmentation approaches on the same dataset: a U-Net architecture for pixel-wise semantic segmentation, Mask R-CNN for instance-level segmentation with region proposals, and DeepLabv3 for atrous-convolution-based dense prediction. A multi-stage variant was also built, splitting the task into background removal followed by anatomical part segmentation, to test whether decomposing the problem improved boundary accuracy on visually complex, low-contrast imagery.

Architecture

Each architecture was evaluated on a held-out test split across the three target classes, comparing segmentation quality and failure modes between the single-stage and multi-stage pipelines to identify which approach generalised better to the dataset's variable lighting, occlusion, and background conditions.

Technology Stack

Business Outcome

The resulting pipeline is a step toward reducing the manual annotation burden in sea turtle photo-ID workflows, aiming to let conservation researchers process substantially more field imagery than manual review allows, freeing analyst time for the judgment calls that still require expert review.

Lessons Learned

Comparing single-stage and multi-stage segmentation on the same dataset made it clear that decomposing a hard vision problem into simpler sub-problems, background removal then part segmentation, often trades training complexity for more interpretable failure modes, which matters when the end users are domain experts who need to trust and sanity-check the output.

Deployment

Research codebase (Jupyter notebooks), not deployed as a production service. Code and methodology are open on GitHub.