Enhancing Gaze Reasoning in Vision Foundation Models for Gaze Following
Vision foundation models struggle with gaze reasoning because they rely too much on what looks semantically important rather than actual gaze cues, leading to poor performance when targets aren't obvious. The authors introduced a head-conditioned local LoRA and an out-of-cone penalty to fix this. This method improves accuracy by allowing the model to focus on true gaze signals instead of relying on visual shortcuts.