← Projects
Ichthyology · Segmentation · Conservation

AI Fish
Counting

We develop computer vision models to automatically detect, segment and count fish. Trained on regional data, they aim to build autonomous ichthyological monitoring systems for freshwater environments.

In development
The starting point

Can AI count fish reliably?

Fish counting is a central problem in ichthyology, fisheries management and freshwater conservation. Today it is usually done manually: a person in front of a screen, counting frame by frame. It is slow, expensive and difficult to scale.

We want to know if modern computer vision models can do it autonomously and with sufficient accuracy for real conservation and fisheries applications. This project exists to answer that question with data, not intuition.

Technical Challenges

Harder than it looks

Counting fish in video remains an open problem. Each environment introduces conditions that generic models handle poorly.

Occlusion

Fish constantly overlap. A fish partially hidden by another generates false negatives or duplicate counts. Instance segmentation must resolve which pixels belong to which individual even when they share visual space.

Camera angles

In confined spaces — channels, indoor pools, inspection chambers — the camera rarely has a clean overhead view. Oblique angles distort fish shape and hinder both detection and species classification.

Confined indoor spaces

Indoor pools and passage ponds create high-density situations: many fish in little space, rapid movement, surface reflections, and variable lighting conditions. None of this appears in standard datasets.

Visual variability

The same fish looks different depending on water turbidity, lighting, angle and stage of its migratory cycle. A model trained under laboratory conditions fails in the field. We are building datasets from real-world conditions.

What we are building

Detection, segmentation and classification

01

Detection

Identify the presence and position of each fish in the frame, including partially occluded individuals.

02

Segmentation

Delineate the exact contour of each individual to separate overlapping fish and obtain accurate counts.

03

Classification

Identify the species of each detected individual, building models trained on regional freshwater species.

Implications

Why counting fish matters

Automatic fish counting has direct applications in fisheries management, biodiversity monitoring and river conservation. Knowing how many individuals of which species pass through a point in a given period remains scarce information, and is rarely obtained continuously and reliably.

Fish lifts, ladders, passage channels and monitoring stations are environments where an autonomous and accurate system could transform the quality of data available for conservation. We are building that technical capability.

Current status

We are working on this

We already have video, annotation data in progress and models in training. Current models perform well in low-density scenes with moderate occlusion. Complex cases — high density, difficult angles and severe occlusion — remain the primary focus of the work.

We do not have a finished product. We have a capability under construction, real data and clarity about the problems that remain to be solved.

Do you work in fish monitoring or fisheries?

If you are facing a counting, identification or ichthyological monitoring problem, let's talk.

Contact Us