HAICK26 Datathon

The challenge

HAICK26 ranked teams across several challenges. Mine was "A Word is Worth a Thousand Pixels": given a photo and a phrase such as "the red car on the left", produce the mask of exactly that object.

My contribution

Rather than fine-tuning, I combined pretrained models without training them. A DINO detector finds the object the phrase describes; I ran it on several rewordings of each phrase and kept the box it was most sure of. SAM then cuts the object's mask from that box. GPU limits on Kaggle's free tier kept the strongest published method out of reach.

The outcome

On this challenge alone we placed fourth. Across the datathon, our team, The Losers, finished first in the overall ranking.