Research map / Video / Video segmentation and tracking
EyeGraph: Modularity-aware Spatio Temporal Graph Clustering for Continuous Event-based Eye Tracking
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces EyeGraph, a 40-participant wearable event-camera eye-tracking dataset, and an unsupervised topology-aware spatio-temporal graph clustering (VGAE with modularity objective) benchmark for pupil tracking.
Problem
Event-based eye tracking relies on supervised learning with RGB guidance or framed event representations, suffers from label sparsity, and lacks datasets reflecting real-world conditions such as varying illuminance and head/body mobility.
Approach
Represents event streams as dynamic spatio-temporal graphs with a GMM-based dynamic threshold edge construction and Hawkes-based edge features, then clusters nodes with a variational graph autoencoder trained with a modularity term, and estimates pupil center-line with tube-model RANSAC.
Key results
- On EyeGraph, the proposed clustering reaches modularity 69.34 and silhouette 0.66, outperforming k-means, DBSCAN, GAE and DGI on modularity, with DMoN a close competitor.
- On 3ET+, EyeGraph achieves p10 of 91.45%, about 8% below the best supervised method (99.42% MambaPupil) but above unsupervised kmeans and DMoN baselines.
- Under varying lighting and mobility conditions, EyeGraph reaches p10 93.67%, about 2% below supervised MambaPupil (95.88%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed dynamic graph construction instead of conventional edge construction | EyeGraph · EBV-Eye | Mo: 72.45 -> 75.70 | Helped |
| Use the proposed topology-aware clustering instead of the strongest baseline clustering method | EyeGraph · EBV-Eye | Mo: 70.34 -> 75.70 | Helped |
| Use the proposed topology-aware clustering instead of DMoN | EyeGraph · EyeGraph | Mo: 68.80 -> 69.34 | Helped |
| Use EyeGraph instead of the best supervised method | pupil coordinate estimation · 3ET+ | p10: 99.42 -> 91.45 | Hurt |
| Add the proposed clustering approach to DMoN (ablation) | DMoN · 3ET+ | p10: +14% | Helped |
| Use EyeGraph instead of the best supervised method | pupil coordinate estimation · EyeGraph (varying lighting and mobility) | p10: 95.88 -> 93.67 | Hurt |
Surprising findings
- Unsupervised EyeGraph approaches supervised methods' p10 accuracy closely, despite no labels.
- Adding the proposed clustering improves DMoN's p10 by about 14%.
Limitations
- Only one eye is tracked monocularly, so gaze features such as saccades and fixations cannot be exploited and binocular data is not available.
- Assumes conjugate eye movements, which may not hold given eye dominance differences.
- Stimulus is a simple moving white circle on screen rather than natural in-the-wild tasks.
Experimental setup
Datasets: EyeGraph EBV-Eye 3ET+
Benchmarks: EyeGraph EBV-Eye 3ET+
Metrics: silhouette coefficient Davies-Bouldin score modularity conductance p10 p5 p1 l2 distance l1 distance
Baselines: kmeans Affinity Propagation Meanshift Spectral Clustering DBSCAN SBM GAE DGI DMoN MambaPupil FreeEVs bigBrains
Builds on: variational graph autoencoder modularity maximization DMoN GSCEventMod Hawkes process
Closest papers by meaning
- DiffEye: Diffusion-Based Continuous Eye-Tracking Data Generation Conditioned on Natural Images (NeurIPS 2025 · similarity 0.86)
- EventZoom: A Progressive Approach to Event-Based Data Augmentation for Enhanced Neuromorphic Vision (AAAI 2025 · similarity 0.85)
- ST-DiffEye: Diffusion-based Continuous Gaze Generation via Joint Scanpath-Trajectory Modeling (NeurIPS 2026 · similarity 0.84)
- GG-SSMs: Graph-Generating State Space Models (CVPR 2025 · similarity 0.84)
- Toward Semantic Gaze Target Detection (NeurIPS 2024 · similarity 0.84)
- EGSST: Event-based Graph Spatiotemporal Sensitive Transformer for Object Detection (NeurIPS 2024 · similarity 0.84)
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.