Gaze Target Estimation Anywhere with Concepts
The paper frames gaze prediction as a promptable task: name or point to a person, then estimate where they are looking.
The authors introduce Promptable Gaze Target Estimation, replacing multi-stage gaze pipelines that depend on head boxes or pose inputs. Their Gaze-Co benchmark contains 120K prompt-annotated image pairs. GazeAnywhere combines subject localization, in/out-of-frame judgment, and gaze heatmap prediction in one model. The paper reports state-of-the-art results across PGE benchmarks, including an out-of-domain clinical dataset. HF Daily Papers' note
The authors introduce Promptable Gaze Target Estimation, replacing multi-stage gaze pipelines that depend on head boxes or pose inputs. Their Gaze-Co benchmark contains 120K prompt-annotated image pairs. GazeAnywhere combines subject localization, in/out-of-frame judgment, and gaze heatmap prediction in one model. The paper reports state-of-the-art results across PGE benchmarks, including an out-of-domain clinical dataset. HF Daily Papers' note
score 4