Sunday, 17 March 2019

$50,000 prize money to find video solution to ID species of fish caught.

Goals of competition:

1. Count fishes and mark frames with current fish number 2. Detect type of each fish 3. Find length of each fish

There was around 1000000 frames for train and test videos to process. To solve this problem we use set of Neural networks: UNET for object localization and ResNet50, DenseNet121, Inception v3 for object classification. Also XGBoost and some heuristic algorithms were used to find sequence of fishes in video.

Problem description:

Sustainable fishing means tracking every fish caught. New tools using automated video processing and artificial intelligence can help responsible fisheries comply with regulations, save time, and lower the safety risk and cost from an auditor on board.

"The winners were so accurate in counting and identifying fish they proved that automated review is now. Not in five years. Now all of the video-review companies are investing in machine learning."
— Christopher McGuire of The Nature Conservancy


Keeping fish populations at healthy levels is critical to sustainable fishing practices, but also puts added requirements on boats to demonstrate they’re meeting the standards. Electronic monitoring systems, like video cameras, could offer an affordable way for fishermen to show their work and keep consumers and fisheries managers confident in the sustainability of their seafood.

The Solution

Data scientists from around the world competed to bring advances in computer vision to count, measure, and identify the species of fish from on-board footage. This challenge was the first to use actual fishing video to develop machine learning tools that can classify multiple parameters, such as size and species.

The Results

The winning algorithm alone achieved above 90% identification accuracy across 5 of 7 species, and 99% accuracy for three species. Meanwhile, predicted counts were within just 1 fish – on an average of 44 fish per video – 83% of the time, and the average error in length was under 2%. One of the top competitors even created a video showing the use of his algorithm in action!

Further explanation of the winner's work here:

With more fisheries adopting video monitoring, the results of this competition have real-world impacts for ocean conservation groups, companies, and fishing communities.