Vision-language trackers are designed to use textual descriptions as an aid, but a new arXiv preprint argues that simply accepting text input is not the same as using it. As visual evidence deteriorates, the authors ask whether the supplied language actually changes tracking decisions—or whether the model falls back on purely visual cues.
The paper, titled “DISTRACT: Degradation-Indexed Study of Textual Reference Across Competing Targets,” introduces a method that indexes different levels of visual degradation and measures how much textual information matters when multiple targets compete for attention. This setup is intended to isolate when and whether language shifts the tracker's output.
The authors report that the phenomenon of accepting but not using text is a real concern, particularly under poor visual conditions. The study is a reminder that a system's ability to ingest multimodal input does not guarantee it exploits that input meaningfully. As a preprint, the work has not yet undergone peer review, and further details of the methodology and results are not covered in the available abstract.