Research map / Video / Video segmentation and tracking
Thermo-JEPA: Learning a Geometry-Grounded Thermal World Model via Cross-Modal Privileged Masking
Biwen Yang, Jin Zhang, Zhe Cao, Ruiheng Zhang
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 RGB-derived confidence-gated topology prior to the teacher attention (ablation) | V-JEPA2 on thermal video · RGBT-World zero-shot action anticipation | Top-1 accuracy (%): 18.6 -> 45.8 | Improved |
| add confidence gating to the structural prior (ablation) | Thermo-JEPA · RGBT-World ablation | Top-1 accuracy (%): 38.5 -> 45.8 | Improved |
| add stop-gradient on teacher targets (ablation) | Thermo-JEPA · RGBT-World ablation | Top-1 accuracy (%): 31.2 -> 45.8 | Improved |
| use DINOv2 semantic topology instead of Sobel edge prior (ablation) | Thermo-JEPA · RGBT-World ablation | Top-1 accuracy (%): 28.7 -> 45.8 | Improved |
| use random topology prior instead of DINOv2 topology (ablation) | Thermo-JEPA · RGBT-World ablation | Top-1 accuracy (%): 18.6 -> 12.4 | Hurt |
| Thermo-JEPA prior vs vanilla V-JEPA2 on spatial probing (ablation) | Thermo-JEPA · RGBT-World diagnostic probe | Spatial mIoU: 32.1 -> 68.7 | Improved |
| Thermo-JEPA vs vanilla V-JEPA2 on trajectory forecasting (ablation) | Thermo-JEPA · RGBT-World diagnostic probe | Trajectory ADE: 28.5 -> 8.4 | Improved |
| Thermo-JEPA vs large generic video foundation models, IR-only probe | Thermo-JEPA (300M) · RGBT-World zero-shot action anticipation | Top-1 accuracy (%): 33.5 -> 45.8 | Improved |
| dense contrastive alignment of IR to RGB instead of privileged prior | V-JEPA2 pretraining · RGBT-World zero-shot action anticipation | Top-1 accuracy (%): 18.6 -> 24.2 vs 45.8 for Thermo-JEPA | Mixed |
Closest papers by meaning
- ThermoGS: Decoupling Physical Surface Attributes for Spatio-Temporal Thermal Field Emulation via 4D Gaussian Splatting (ECCV 2026 · similarity 0.88)
- TherA: Thermal-Aware Visual-Language Prompting for Controllable RGB-to-Thermal Infrared Translation (CVPR 2026 · similarity 0.88)
- ThermalGaussian: Thermal 3D Gaussian Splatting (ICLR 2025 · similarity 0.87)
- ThermalGen: Style-Disentangled Flow-Based Generative Models for RGB-to-Thermal Image Translation (NeurIPS 2025 · similarity 0.86)
- DarkVGGT: Seeing Through Darkness Using Thermal Geometry without Daylight Tax (NeurIPS 2026 · similarity 0.85)
- Video World Models with Long-term Spatial Memory (NeurIPS 2025 · 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.
