DarkVGGT enables 3D reconstruction in darkness using thermal cues without daylight tax
New feed-forward framework uses thermal imaging for robust 3D estimation in low-light.
Deep Dive
Researchers introduce DarkVGGT, an RGB-T feed-forward geometry framework that uses physics-aware thermal modeling for 3D reconstruction in dark scenes. It features thermal factorization to extract geometry-consistent cues and geometry-shared thermal routing to inject structural guidance into the RGB stream, preserving performance in well-lit environments. Consistent improvements on low-visibility RGB-T benchmarks for depth and pose estimation.
Key Points
- Physics-aware thermal factorization extracts emissive-dominant, geometry-consistent cues while isolating sparse reflective residuals.
- Geometry-shared thermal routing injects reliability-aware structural guidance into the RGB stream, preserving performance in well-lit conditions.
- Outperforms existing feed-forward methods on low-visibility RGB-T benchmarks for depth and camera pose estimation.
Why It Matters
Enables reliable 3D perception in darkness for autonomous vehicles and robotics without daylight performance loss.