/Seagulls

Hackathon challenge ML Hackathon: Utashud's wild life!

Primary LanguageJupyter NotebookMIT LicenseMIT

Hackathon challenge ML Hackathon: Utashud's wild life!

Code Size

Hackathon Challenge: Seagulls

Utashud Island, part of the South Kamchatka Federal Reserve, lies 4 kilometers from the Pacific coast of Kamchatka in Vestnik Bay. Sea otters, anthurs, spotted seals, about 50,000 birds of ten different species live on it, including a colony of 10,000 hatchet pairs. At the hackathon, we are interested in two inhabitants of the island - spotted seal and techno-ocean gull, which are monitored through camera traps in the summer season.

On Utashuda there is a reproductive rookery of spotted seals and one of the largest settlements of the slaty-backed gull in Kamchatka (4 thousand pairs). For 1 season of observing these animals, more than 1800 photographs of various quality can be accumulated. Through these photographs, inspectors assess the welfare and abundance of animals.

Develop an algorithm to count seagulls that are larger than 5x5 pixels in a photograph and help the researchers.

Task metric - (1-RMSE). You can read about how it is calculated in the task baseline.

Used technology stack:

Decision progress

Dataset extension

The initial dataset consisted of 500 photos for the training sample, and 99 photos for the validation sample. I decided to expand the training dataset using the albumentations framework, which resulted in 1198 images for training.

The code for expanding the training sample can be viewed in this folder.

Choose model

For training, I decided to take the yolov5x model with an image size of 640 * 640, the maximum possible batch size (and this is -1 which automatically fills free space in RAM), 128 epochs.

Submition score and place that turned out during the hackathon

Public -1.4052026941611389

Private -1.126590712784548

I took 5th place in the top 10