PelAtlas

Context

Capture–recapture studies rely on consistent individual identification across repeated encounters. In marine mammal research, this is often done via photo-identification (photo-ID), using stable visual cues such as dorsal fin shape, nicks, scars, and pigmentation patterns.

In practice, the limiting factor is rarely the downstream ecological modeling — it’s the manual work needed to turn raw survey imagery into a traceable catalog: filtering images for animal presence, isolating identity-relevant regions, matching against a growing set of known individuals, and keeping the catalog consistent over time.

PelAtlas is motivated by a long-term, real-world dataset context: a large archive of boat-survey imagery accumulated over ~20 years, with heterogeneous quality and incomplete metadata, where only a subset of images is suitable for identification.

Project focus

PelAtlas aims to build a maintainable pipeline that reduces expert workload while preserving scientific traceability.

The project is structured as a stage-wise workflow aligned with the practical bottlenecks of photo-ID:

• Stage 1: Presence filtering (high recall): separate likely relevant images from background.
• Stage 2: Region proposals: extract typed identity evidence (fin-first; multi-region later).
• Stage 3: Embeddings + within-sighting grouping: aggregate evidence across multiple frames to stabilize matching.
• Stage 4: Retrieval + open-set handling + expert verification: propose top-k candidates with an explicit “unknown” option, then log human decisions as the scientific record and as training signal.

A central design principle is decision support rather than full automation: the system proposes evidence and candidates; experts confirm, reject, or create a new identity entry.

Collaboration

PelAtlas is developed in a small, tightly coupled collaboration:

• Oihana Olhasque — marine mammal photo-ID workflow context, field and expert perspective, requirements shaping
• Jan Meischner — workflow and system design, ML pipeline architecture, evaluation design, implementation coordination

The project is motivated by an NGO field-work context and is designed around real operational constraints rather than benchmark-style assumptions.

My role

I lead the technical side: defining the stage-wise pipeline, designing the data model and provenance strategy, and building the ML-assisted workflows that connect ingestion, model inference, review, and catalog updates.

A key part of my role is ensuring that the system remains scientifically robust as it scales: defining clear baselines, structuring evaluation across stages, and making sure that human decisions remain traceable and reproducible over time.

Implementation

Implementation is approached as a minimal, maintainable system that supports both historical backfill and continuous ingestion:

• survey and sighting as first-class organizational units, enabling within-sighting grouping and context-aware processing
• asynchronous processing for compute-heavy steps (ingest, inference, embedding, indexing)
• traceable decision logging: for each verification, store the evidence shown, the model version, and the final human outcome
• open-set behavior as a hard requirement to avoid systematic incorrect merges in the catalog

Stack

Expected outcomes

• A reproducible, continuously ingestible workflow from raw survey images to verified catalog updates
• Baselines and ablations for each stage, plus end-to-end workflow evaluation (expert effort and catalog quality)
• A fin-first re-identification setup for bottlenose dolphins (retrieval-based, open-set, auditable)
• A path to multi-region evidence (Risso-ready) without changing the overall retrieval framing
• A structured foundation for scaling label efficiency via active learning (and, later, carefully scoped non-expert tasks)

Learnings

• For long-term datasets, “data and workflow engineering” is not a side task — it often determines what modeling is feasible at all.
• Open-set handling is not optional in photo-ID: forced matches can silently corrupt capture histories.
• Within-sighting aggregation is not just an optimization; it aligns the system with how experts reason and reduces redundant decisions.