From Fishnet to Meenu: from detector accuracy to reviewable fisheries evidence
In brief
- Fishnet is an external open dataset; Meenu is a systems problem around event evidence and review.
- Random frame splits do not establish cross-vessel generalisation.
- The operational metric is validated evidence per unit review effort, with an audit of what automation misses.
Species detection is only one component of electronic monitoring. A useful monitoring system must also reconstruct events, preserve evidence provenance, quantify uncertainty and support human verification.
Dataset provenance first
The Fishnet Open Images Database was introduced by Justin Kay and Matt Merrifield. The 2021 paper describes 86,029 images across 34 object classes and discusses species similarity, class imbalance and difficult deck conditions. Fishnet is an external research resource, not a dataset I created.
The dataset is useful precisely because it exposes domain-specific computer-vision challenges. But an open benchmark does not prove that a deployed model generalises across vessels, camera geometries or fisheries.
The unit of work is an event
One catch can appear in many frames, disappear behind a crew member and reappear at a different orientation. Therefore:
frame detections → track/object → catch interaction → reviewed event.
Detection and segmentation provide observations. Tracking/temporal aggregation forms the event. Human review resolves ambiguity and records corrections.
The event record should retain the original evidence, model/version, prediction, confidence, reviewer outcome and any later correction.
Generalisation
For deployment I propose to use:
- vessel-held-out testing;
- trip/time-held-out testing;
- species-stratified results;
- camera-condition slices;
- calibration / selective prediction;
- explicit sister-species confusion.
This is consistent with the wider trajectory of deep-learning evaluation: performance under distribution shift and the reliability of confidence matter as much as average in-distribution accuracy.
Human review as part of the model
A useful AI-assisted review system needs at least two queues:
- priority queue — high uncertainty, rare/ETP classes, disagreement and model-shift signals;
- random audit queue — unbiased footage/events, including material the model did not flag.
Without the second queue, the system cannot estimate false negatives introduced by automated filtering.
What “impact” should mean
I propose to compare assisted vs unassisted review on the same evaluation set and report:
- review minutes per hour of footage;
- event recall;
- correction burden;
- unresolved events;
- random-audit misses;
- cross-vessel performance;
- time from capture to validated report.
A lower review time is meaningful only when evidence quality and coverage remain comparable or improve.
Public context
The Nature Conservancy and Newlab publicly selected Blurgs.AI for a Global Ocean Innovation Challenge pilot focused on electronic-monitoring training and analytics. That is a pilot-selection milestone, not evidence of completed conservation impact. Any impact claim should come from a transparent deployment evaluation.
The research connection is direct: representation learning determines which distinctions are recoverable; uncertainty determines when the system should defer; provenance and human correction determine whether the resulting evidence can be challenged and revised.