Researchers have introduced DISTRACT, a new benchmark designed to study how textual references influence vision-language trackers (VLTs), especially as visual evidence degrades. The benchmark systematically tests four text priors against four levels of image degradation, holding video and target elements constant. Evaluations of five VLTs revealed that Molmo2 and JointNLT demonstrate directed text use, with some effects persisting even under degradation, while UVLTrack, MMTrack, and DUTrack showed weaker responses to semantic text changes. AI
IMPACT This research provides a new method for evaluating the robustness of vision-language models to degraded visual input and textual prompts.
RANK_REASON The item is a research paper detailing a new benchmark and evaluation of existing models. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
