More and more companies, as well as government agencies and researchers, are using computer vision to review video data for stock assessments and electronic monitoring of fishing vessel trips.
Computer vision (CV) is an artificial intelligence (AI) technology that has eliminated countless hours of manual review of often empty video and provided more accurate analysis of existing data. But CV comes with some limitations, the most limiting of which is the need for extensive training data. “The challenge with computer vision models is that they need a lot of training data; hundreds of thousands of images,” said Alexander Dungate, co-founder of the Vancouver-based OnDeck AI.
Working together, Dungate and OnDeck AI co-founder Sepand Dyanatkar are using a different model to analyze video. “We’re using Vision Language Models (VLM),” said Dungate. “Unlike traditional computer vision, our models can identify what you’re looking for without all the training data. Computer vision basically memorizes the training data while VLM learns to reason, almost like a human, based on just a few images.”
Dungate and Dyanatkar founded OnDeck AI in 2025 and have received several million dollars in funding and grants to pursue their ambitious project. If successful, VLM models will take the use of artificial intelligence in fisheries science and management to the next level. But there are still hurdles to overcome.
“We’re working with some First Nations fishers in the halibut, blackcod, and rockfish longline fishery around Vancouver,” said Dungate. “And we’re spinning up a project to develop an electronic monitoring program for the lake fisheries in northern Manitoba. But so far, we don’t have systems fully deployed and operational in any fisheries.”
While Dungate noted some regulatory limitations to VLM deployment, somewhat like those placed on self-driving cars, as well as financial constraints, he acknowledged that there are technological obstacles.
“There is something called a context window,” he explained. “Which is the limit on how much data we can upload and process. It’s expanding, but not enough, so what we do is split the data to create more context windows so that we can handle the volume while maintaining high performance and accuracy.”
Dungate said that he believes the technology could be deployed on fishing vessels in three to five years if there is enough investment to make it happen. “Right now, I’d say we’re at about 40 percent of what we need,” he said. “What it’s going to require is a commitment of millions of dollars over several years.
“We’ve published a paper on this,” said Dungate. “We’ve beaten the benchmarks for species identification, and we have a program deployed in other private industries, but getting this into fisheries monitoring and conservation is what we are most passionate about.”
