The authors sampled representative keyframes from each video and extracted two forms of attention from a frozen large vision-language model.
Learned attention adaptation improves language-guided video object segmentation
ATLAS converts a frozen vision-language model’s spatial and temporal attention patterns into localization maps that guide SAM2.
Research Lab
Hamidreza Dastmalchi · Ali Cheraghian · Aijun An · Abdelwahed Khamis · Lars Petersson
York University · Macquarie University · CSIRO Data61
Research Digest··3 min read
Thread:Visually Grounded Reasoning
Dastmalchi and colleagues introduce ATLAS, a supervised adapter for segmenting video objects described by natural-language expressions.
§