Research map / Label-efficient and robust learning / Domain adaptation and generalization
Rapidly Deploying On-Device Eye Tracking by Distilling Visual Foundation Models
Cheng Jiang, Jogendra Nath Kundu, David Colmenares, Fengting Yang, Joseph Robinson, Yatong An, Ali Behrooz
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune DINOv3 on synthetic data instead of linear probing | DINOv3 ViT-B · Project Aria ET (2,222 participants) | E50U50: 5.47 -> 2.01 | Improved |
| add self-distillation and unlabeled real data to synthetic fine-tuning (optimized VFM) | DINOv3 ViT-B · Project Aria ET (2,222 participants) | E50U50: 2.01 -> 1.33 | Improved |
| same, at the tail | DINOv3 ViT-B · Project Aria ET (2,222 participants) | E90U90: 12.11 -> 8.32 | Improved |
| use DARE-GRAM domain adaptation instead of the proposed teacher optimization | DINOv3 ViT-B · Project Aria ET (2,222 participants) | E50U50: 2.01 -> 4.84 | Hurt |
| distill into on-device model with DistillGaze instead of synthetic supervision | FBNet 256K · Project Aria ET (2,222 participants) | E50U50: 3.48 -> 1.44 | Improved |
| same, at the tail | FBNet 256K · Project Aria ET (2,222 participants) | E90U90: 14.84 -> 8.45 | Improved |
| add self-distillation with unlabeled real data to on-device model (no VFM) | FBNet 256K · Project Aria ET (2,222 participants) | E50U50: 3.48 -> 2.20 | Improved |
| add optimized VFM teacher on top of on-device self-distillation | FBNet 256K · Project Aria ET (2,222 participants) | E50U50: 2.20 -> 1.44 | Improved |
| use MSE regression loss instead of DINO loss for teacher (ablation) | Optimized VFM · Project Aria ET (2,222 participants) | E50U50: 1.50 -> 1.33 | Improved |
| add synthetic supervision to self-distillation and pseudo-labels (ablation) | Optimized VFM · Project Aria ET (2,222 participants) (loss weight ablation) | E50U50: 9.53 -> 1.33 | Improved |
| add loss weight scheduler (ablation) | Optimized VFM · Project Aria ET (2,222 participants) (loss weight ablation) | E50U50: 8.73 -> 1.33 | Improved |
| add self-distillation and pseudo-label losses to synthetic supervision (ablation) | Optimized VFM · Project Aria ET (2,222 participants) (loss weight ablation) | E50U50: 2.22 -> 1.33 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DARE-GRAM — DARE-GRAM: Unsupervised Domain Adaptation Regression by Aligning Inverse Gram Matrices (CVPR 2023) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- EV-Eye: Rethinking High-frequency Eye Tracking through the Lenses of Event Cameras (NeurIPS 2023 · similarity 0.88)
- Towards End-to-end Video-based Eye-Tracking (ECCV 2020 · similarity 0.88)
- OmniGaze: Reward-inspired Generalizable Gaze Estimation in the Wild (NeurIPS 2025 · similarity 0.85)
- ST-DiffEye: Diffusion-based Continuous Gaze Generation via Joint Scanpath-Trajectory Modeling (NeurIPS 2026 · similarity 0.85)
- DiffEye: Diffusion-Based Continuous Eye-Tracking Data Generation Conditioned on Natural Images (NeurIPS 2025 · similarity 0.85)
- Enhancing 3D Gaze Estimation in the Wild using Weak Supervision with Gaze Following Labels (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
