Human gaze does more than point to visual targets; it serves as a subtle indicator of social intent within static images, whereas standard models typically process individuals independently, treating gaze as an i.i.d. quantity or predicting social semantics in isolation. Recent multi-person methods attempt to address this but often treat social relations as rigid, post-hoc classifications decoupled from the gaze estimation process. This oversimplification fails to capture the nuanced nature of social intent, which acts as an underlying driver of gaze behavior rather than a secondary categorical output. We address these limitations by proposing ANCHOR, a target-centric paradigm designed to decode gaze-anchored social intent by modeling the joint distribution of visual attention and latent implicit relations. Our approach surfaces these dependencies as the latent structural scaffolding of gaze behavior. The architecture utilizes a relational attention mechanism to capture fine-grained interper- sonal links, leveraging feature-wise modulation for efficient multi-person parsing from a single vision backbone. To stabilize the training of this coupled formulation, we implement an optimization synergy to resolve the inherent conflicts between spatial gaze accuracy and latent social reasoning soning. This approach ensures robust generalization by seeking stable, flat minima while simultaneously harmonizing competing task gradients. We validate our framework on an extended benchmark featuring dense multi-person annotations and novel social influence rankings. Our results demonstrate state-of-the-art performance and provide the first quantitative evidence that implicit social hierarchies can be robustly disentangled and learned directly from static gaze patterns.

Scanpath prediction for egocentric videos is a challenging task because human gaze is constantly moving with ongoing actions, while the first-person viewpoint is also continuously changing. Existing approaches incorporate additional inputs, such as hand movements, body poses, or audio signals, which may not always be available and can reduce generalization. In this work, we present VSP, a Visual Scanpath Predictor for egocentric video that predicts the next fixation location using only past RGB video frames and fixation history, without needing additional modalities. Based on the USP architecture and evaluated on the EGTEA Gaze+ dataset, VSP includes a Dynamic Attention Module (DAM), modality-specific encoders, an Attentive Convolutional LSTM (ALSTM), and a convolutional Gated Multimodal Unit (GMU) to pre- dict the next fixation in a scanpath. Unlike prior work that predicts continuous saliency maps, VSP directly predicts the next fixation location, resulting in predictions that align with real-world human scanpaths. Experiments demonstrate that VSP achieves an F1 score of 55.30%, with ablation studies confirming that each architectural component contributes meaningfully to overall performance. Our results suggest that an accurate prediction of egocentric fixation can be achieved without relying on computationally expensive or environment-specific additional inputs.

The analysis of human gaze behavior serves as a critical, objective biomarker in pediatric research, particularly for assessing neuro-developmental disorders and evaluating visual rehabilitation in unstructured, naturalistic settings. While unconstrained, single-camera video recordings offer a non-invasive alternative to eye-tracking hardware, they inherently collect highly sensitive clinical data from vulnerable minors, raising severe ethical and privacy concerns. This lack of secure data-sharing methodologies represents a significant bottleneck for collaborative research. Traditional face anonymization techniques (blurring, pixelation) are inadequate, as they distort key facial features and render subsequent gaze behavior analysis impossible. To bridge this gap, this paper introduces a novel, training-free pipeline that uniquely integrates a state-of-the-art latent diffusion inversion model, NullFace, with an advanced scene-aware gaze estimator, Gaze-LLE. By leveraging diffusion-based identity suppression, our approach precisely anonymizes the subject’s face while preserving the subtle ocular and geometric cues required for high-fidelity gaze analysis. We validate our pipeline on the highly dynamic and unconstrained ChildPlay-Gaze dataset, which serves as a rigorous proxy for naturalistic and unstructured clinical environments By quantifying the gaze deviation introduced by anonymization and demonstrating an optimal balance between identity suppression and gaze preservation, we show that our architecture successfully decouples identity from gaze metrics. This work establishes the first viable methodology for privacy-preserving pediatric gaze estimation, paving the way for secure, large-scale dataset sharing in neurodevelopmental and rehabilitative research.

Head-mounted eye tracking is pivotal for extended reality but suffers from mechanical slippage. Unlike model-based approaches, regression-based Pupil Center-Corneal Reflection (PCCR) methods are not very robust against sensor displacement as they discard absolute geometric information of the eye features. This paper presents PCCR+ , a dynamic correction method that leverages absolute pupil dynamics to mitigate slippage error. A correction term is introduced, derived from the discrepancy between absolute pupil estimates and relative PCCR estimates. Utilizing a secondary "slip-calibration", a mapping is established between the PCCR vector and a dynamic gain coefficient. Simulation evaluations demonstrate that PCCR+ reduces mean absolute error by approximately 82% compared to the conventional PCCR method. While empirical validation remains for future work, this theoretical framework establishes a robust mathematical foundation for mitigating slippage errors in off-the-shelf solutions, without necessitating the complex geometric reconstruction required by model-based approaches.