Salmon Counting · Episode 6 of 8 · 49 min
The Echogram: Counting Salmon With a Single Image
A cheap computer at a roadless field camp can count fish if hundreds of sonar frames are collapsed into one picture, an echogram, which a lightweight neural network reads in a single pass. The episode explains the echogram, a ResNet-18 model and the Caltech Fish Counting dataset from the Kenai. The trade-off is more error than heavy models, in exchange for working where there is no compute and no internet.
Who's who
- Kai Van BruntMIT co-author of the echogram counting paper
- Justin KayMIT co-author of the paper and the Caltech dataset
- Timm HauckeMIT co-author of the paper
- Pietro PeronaCaltech researcher and co-author
- Grant Van HornUMass Amherst researcher and co-author
- Sara BeeryMIT researcher and co-author
Where it happens
- Kenai River, Alaska+1 more episodeSource of footage in the Caltech Fish Counting dataset
Timeline
- 2018-2021Agencies increasingly adopt imaging sonar for turbid rivers.
- July 2022Caltech Fish Counting dataset released at ECCV 2022.
- 2022-2024Detector-tracker pipelines become the accuracy benchmark but are impractical at remote camps.
- February 2025Paper on counting fish with temporal representations of sonar video is posted.
Still debated
The echogram method shows 23 to 31 percent error, which may be tolerable for a rough in-season trend but inadequate for precise stock assessment. The work demonstrates feasibility rather than readiness for regulatory use.