Thousands of images per season
Large marine mammal surveys accumulate image collections faster than they can be filtered, reviewed, and linked back to survey context.
Turning years of dolphin survey imagery into reliable research data
PelAtlas structures and scales marine mammal photo-identification workflows, built around structured human annotation and designed to evolve toward ML-assisted decision support.
© Association AL LARK
Why photo-identification doesn't scale
Photo-identification workflows break down at scale.
Large-scale marine mammal surveys generate more imagery than can be consistently filtered, reviewed, and integrated into long-term catalogs using manual workflows.
Manual filtering and identification
Researchers spend substantial time separating useful imagery from noise and comparing individuals against existing catalogs.
Maintaining consistent catalogs
Long-term identity decisions need provenance, repeatable review, and careful handling of uncertainty across seasons and teams.
A structured, scalable workflow for marine photo-identification
PelAtlas replaces fragmented processing with maintainable research infrastructure that supports auditability, long-term catalog maintenance, and future model-assisted workflows.
Structured workflows instead of ad-hoc processing
Move from fragmented file handling to repeatable research operations.
Traceable decisions and provenance
Keep evidence and catalog changes understandable over time.
Human validation remains central
Identity decisions stay reviewable and accountable to researchers.
Designed to integrate ML incrementally
Add assistance features carefully without losing scientific control.
Platform capabilities
Current and planned functionality of the PelAtlas platform
Data foundations
Organizing surveys, imagery, and provenance for durable research datasets.
Survey ingestion
Structured import of surveys, images, sightings, and metadata into a shared research workflow.
Image gallery
Browse, filter, and inspect large survey image collections with dataset context preserved.
Workflow systems
Supporting structured human review, collaboration, and validation.
Annotation workspace
Manual annotation with explicit review states, evidence capture, and traceable outcomes.
Collaborative workflows
Shared queues and team-oriented review flows for structured annotation and validation.
Citizen science integration
Task distribution to broader contributor groups with quality control and validation checkpoints.
ML decision support
Planning for model-assisted triage, retrieval, and active learning.
ML-assisted filtering
Automatic triage of relevant imagery to reduce the initial manual screening workload.
Identity evidence extraction
Detection of identity-relevant regions, such as dorsal fins, to support downstream review.
Candidate-based identification
Retrieval of likely matches to support open-set review instead of forcing a closed-set answer.
Active learning workflows
Model-guided prioritization of the most informative images for efficient annotation effort.
Built around real annotation workflows
PelAtlas is designed around practical review tasks, from dataset triage to detailed annotation and image comparison.
Human-in-the-loop by design
The platform is aligned with scientific workflows where expert review remains central and machine learning is introduced as support rather than opaque automation.
Human review stays authoritative
PelAtlas is built for decision support, with experts confirming matches, unknowns, and uncertain cases.
Annotation is structured, not informal
Reviews happen through explicit workflow steps so evidence, rationale, and outcomes remain auditable.
ML can improve without replacing oversight
As assistance features mature, they can reduce workload while preserving traceability and scientific control.
Conceptual pipeline
ML-assisted identification workflow (in development)
Step 1
Image filtering
Surface relevant survey imagery before deeper review begins.
Step 2
Identity evidence extraction
Locate dorsal-fin evidence and other identity-relevant regions.
Step 3
Candidate retrieval
Compare likely matches without forcing a closed-set answer.
Step 4
Expert validation
Expert review confirms matches, new individuals, and uncertainty.
This process is currently under development
Why this matters
Capture–recapture studies
Reliable identity decisions support population estimation and demographic analysis.
Population monitoring
Structured historical records help track abundance, site fidelity, movement and distribution across seasons and years.
Cross-regional connectivity
Connecting catalogs across institutions enables the study of population connectivity and supports conservation at broader geographic scales.
Reproducible research
Traceable workflows and preserved provenance make methods easier to review, reuse, and trust.
Work with PelAtlas
Whether you're working with long-term survey data or building a new catalog, PelAtlas provides the infrastructure for structured, scalable photo-identification research.