MATTE: Multi-task multi-scale attention

Download PDF
Open Webpage

Gjorgji Strezoski, Nanne van Noord, Marcel Worring

Published: 2023, Last Modified: 11 Dec 2025Comput. Vis. Image Underst. 2023EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Highlights•MATTE is a MTL attention that fits any multi-scale architecture.•MATTE combats destructive interference, a known problem in MTL methods.•We report competitive performance with baselines and qualitative findings.