Research map / 3D vision / Human pose estimation
Cross-Encoder for Unsupervised Gaze Representation Learning
Yunjia Sun, Jiabei Zeng, Shiguang Shan, Xilin Chen
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 |
|---|---|---|---|
| add gaze-similar pair training in addition to eye-consistent pairs (ablation) | Cross-Encoder · UTMultiview 100-shot, w/ head pose | angular error (deg): 10.6 -> 8.0 | Improved |
| add gaze-similar pair training in addition to eye-consistent pairs (ablation) | Cross-Encoder · Columbia 100-shot, w/ head pose | angular error (deg): 7.6 -> 6.6 | Improved |
| add residual loss (ablation) | Cross-Encoder · Columbia 100-shot, w/ head pose | angular error (deg): 6.7 -> 6.6 | Improved |
| add residual loss (ablation) | Cross-Encoder · UTMultiview 100-shot, w/ head pose | angular error (deg): 7.4 -> 8.0 | Hurt |
| add residual loss (ablation) | Cross-Encoder · MPIIGaze 100-shot, w/ head pose | angular error (deg): 7.2 -> 7.5 | Hurt |
| use gaze feature instead of eye feature (ablation) | Cross-Encoder · UTMultiview 100-shot, w/ head pose | angular error (deg): 15.5 -> 8.0 | Improved |
| swap features (Cross-Encoder) instead of equal feature constraint (ablation) | auto-encoder · Columbia 100-shot, w/ head pose | angular error (deg): 9.2 -> 6.6 | Improved |
| swap features (Cross-Encoder) instead of plain auto-encoder (ablation) | auto-encoder · UTMultiview 100-shot, w/ head pose | angular error (deg): 18.0 -> 8.0 | Improved |
| vary gaze feature dimension dg from 9 to 15 (ablation) | Cross-Encoder · Columbia 100-shot, w/ head pose | angular error (deg): 6.7 -> 6.4 | No clear effect |
| increase eye feature dimension from 16 to 64 (ablation) | Cross-Encoder · Columbia 100-shot, w/ head pose | angular error (deg): 6.7 -> 6.5 | No clear effect |
| add more unlabeled data (XGaze, TabletGaze, FreeGaze union) instead of XGaze alone (ablation) | Cross-Encoder cross-dataset · Columbia 100-shot cross-dataset | angular error (deg): 7.76 -> 7.09 | Improved |
| use Cross-Encoder instead of SimCLR/BYOL (ablation) | gaze representation · UTMultiview 50-shot | angular error (deg): 14.4 -> 8.8 | Improved |
Closest papers by meaning
- Unsupervised Gaze Representation Learning from Multi-view Face Images (CVPR 2024 · similarity 0.90)
- Enhancing 3D Gaze Estimation in the Wild using Weak Supervision with Gaze Following Labels (CVPR 2025 · similarity 0.88)
- De^2Gaze: Deformable and Decoupled Representation Learning for 3D Gaze Estimation (CVPR 2025 · similarity 0.87)
- Bootstrap AutoEncoders With Contrastive Paradigm for Self-supervised Gaze Estimation (ICML 2024 · similarity 0.87)
- OmniGaze: Reward-inspired Generalizable Gaze Estimation in the Wild (NeurIPS 2025 · similarity 0.86)
- Gaze-LLE: Gaze Target Estimation via Large-Scale Learned Encoders (CVPR 2025 · similarity 0.86)
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.
