PaGE: Towards Practical Human-Level
Gaze Target Estimation

Practical Gaze Estimator with human-level accuracy and deployable student models

Zhoutong Ye1,*, Chengwen Zhang1,*, Zhaibin Cui1, Mingze Sun1, Jiaqi Liu1, Xiangwu Li2, Qingyang Wan1, Qingyang Wan1, Chang Liu1, Xutong Wang1, Huan-ang Gao1, Yu Mei1, Chun Yu1, Yuanchun Shi1

1Tsinghua University    2Jinan University    *Equal contribution    †Corresponding author

Paper Demo Code Models BibTeX

Abstract

Gaze target estimation, the task of predicting where a person is looking in a scene, is crucial to understanding human attention and intent. It is a challenging task that combines high-level understanding of global scene semantics and precise spatial reasoning using human appearance (e.g. pose, eye orientation). As a result, human-level performance remains elusive for existing models, limiting their practical application. To this end, we propose PaGE (Practical Gaze Estimator), a gaze estimation model that explicitly models the complex interaction between scene and head features. Using a PaGE model with a large ViT-H+ backbone as the teacher, we further distill student models with lighter backbones on a much larger and more diverse unlabeled dataset. The architectural improvements and novel training recipe allow PaGE to achieve state-of-the-art performance on several gaze estimation tasks, outperforming humans in 7 out of 9 metrics while reducing the human–AI gap by at least 60% in the remaining 2. The distilled student models retain most of the teacher's performance while being lightweight enough for practical deployment on robots and consumer devices.

7 / 9
metrics surpass human performance on GazeFollow, VAT & ChildPlay
≥60%
reduction in human–AI gap on the remaining metrics
~10%
of teacher FLOPs for distilled student models with SOTA accuracy

Method

PaGE builds upon DINOv3 with a Scene-head Interaction Module (SIM) that uses cross-attention between scene and head branches to model inter-branch feature interaction in a ViT-native manner. Training follows a two-stage recipe: decoder-only training with a frozen backbone, followed by supervised finetuning (SFT) of the full model. Lightweight student models are trained via token-level feature distillation from a PaGE ViT-H+ teacher on 1.17M unlabeled head crops from MPII and OpenImages V7.

PaGE architecture diagram
Figure 1. Architecture of PaGE. Bottom left: our SIM layer. Bottom right: the token-concat head branch. Ldistill marks where token-level feature distillation is applied during student training.

Results

PaGE ViT-H+ achieves human-level performance on GazeFollow, VideoAttentionTarget (VAT), and ChildPlay. Distilled variants (ViT-S, ViT-S+, ViT-B) retain most of the teacher's capability at 5–10% of its compute cost.

Table 3. Comparison with Prior Art

Model GazeFollow VideoAttentionTarget ChildPlay
AUC↑ Avg L2↓ Min L2↓ AUC↑ L2↓ APin/out↑ AUC↑ L2↓ APin/out↑
Recasens et al. (2015) 0.878 0.190 0.113 — — — — — —
Chong et al. (2018) 0.896 0.187 0.112 0.833 0.171 0.712 — — —
Lian et al. (2019) 0.906 0.145 0.081 — — — — — —
Chong et al. (2020) 0.921 0.137 0.077 0.860 0.134 0.853 — — —
Fang et al. (2021) 0.922 0.124 0.067 0.905 0.108 0.896 — — —
Bao et al. (2022) 0.928 0.122 — 0.885 0.120 0.869 — — —
Jin et al. (2022) 0.920 0.118 0.063 0.900 0.104 0.895 — — —
Gupta et al. (2022) 0.943 0.114 0.056 0.914 0.110 0.879 0.919 0.113 0.983
Miao et al. (2023) 0.934 0.123 0.065 0.917 0.109 0.908 — — —
Tafasca et al. (2023) 0.939 0.122 0.062 0.914 0.109 0.834 0.935 0.107 0.986
Tafasca et al. (2024) 0.944 0.113 0.057 — 0.107 0.891 — 0.106 0.990
Song et al. (2024) 0.949 0.105 0.047 0.938 0.102 0.905 — — —
Ryan et al. (2025)* 0.958 0.099 0.041 0.937 0.103 0.903 0.951 0.101 0.994
Gemini 3.5 Flash — 0.128 0.067 — 0.138 — — 0.111 —
PaGE ViT-S Distill 0.964 0.086 0.033 0.964 0.074 0.937 0.970 0.075 0.997
PaGE ViT-S+ Distill 0.965 0.086 0.033 0.965 0.074 0.939 0.970 0.075 0.997
PaGE ViT-B Distill 0.966 0.081 0.029 0.969 0.068 0.945 0.973 0.070 0.997
PaGE ViT-H+ 0.966 0.080 0.029 0.972 0.064 0.951 0.975 0.069 0.995
Human 0.924 0.096 0.040 0.921 0.051 0.925 0.911 0.048 0.993

Comparison between prior art, the PaGE family, and humans. All four PaGE models far outperform the previous SOTA, with PaGE ViT-H+ and PaGE ViT-B Distill achieving human-level performance. *Ryan et al. (2025), Gaze-LLE ViT-L, results reported in the original paper.

Qualitative results on benchmark datasets
Figure 2. Qualitative results of PaGE ViT-H+ on GazeFollow, VideoAttentionTarget, ChildPlay, GOO-Real, and Zhang et al. Green boxes mark the target head; gaze lines and heatmaps show model predictions.
Out-of-distribution qualitative results
Figure 3. Out-of-distribution qualitative results. PaGE generalizes beyond human gaze in realistic images to animals, cameras, and animation.
Screen gaze estimation results
Figure 4. Fine-grained screen gaze estimation. Training with screen data during SFT substantially improves gaze localization on monitors.

Model Zoo

We release pretrained checkpoints for the PaGE family. The ViT-H+ teacher is finetuned end-to-end; student models are distilled from the teacher and finetuned on the labeled training set.

Model Backbone GFLOPs Checkpoint
PaGE ViT-H+ DINOv3 ViT-H+ 2373.6 vithplus_ft.pt
PaGE ViT-B Distill DINOv3 ViT-B 283.1 vitb_distill.pt
PaGE ViT-S+ Distill DINOv3 ViT-S+ 115.2 vitsplus_distill.pt
PaGE ViT-S Distill DINOv3 ViT-S 96.9 vits_distill.pt

See the code repository for installation instructions and usage examples.

BibTeX

@misc{ye2026pagepracticalhumanlevelgaze,
      title={PAGE: Towards Practical Human-level Gaze Target Estimation}, 
      author={Zhoutong Ye and Chengwen Zhang and Zhaibin Cui and Mingze Sun and Jiaqi Liu and Xiangwu Li and Qingyang Wan and Chang Liu and Xutong Wang and Huan-ang Gao and Yu Mei and Chun Yu and Yuanchun Shi},
      year={2026},
      eprint={2607.04860},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2607.04860}, 
}