🎉 IROS 2026 Accepted!
Conventional SLAM systems that rely on photometric alignment or point features often fail in the texture-sparse, low-contrast imagery produced by thermal cameras. We present PI-EAT-SLAM (Physics-Informed Edge-Aware Thermal Gaussian Splatting SLAM), the first thermal-depth 3D Gaussian Splatting SLAM framework. Instead of depending on raw photometric intensity, our approach exploits robust geometric edge features to achieve reliable tracking and mapping in visually degraded environments. A Stefan-Boltzmann law-based rescaling module physically enhances thermal contrast without amplifying noise; a multi-stage Gradient-aware Edge-KLT (GE-KLT) tracker establishes highly reliable correspondences via dual-constraint outlier rejection; and an edge-aware smoothness loss preserves sharp thermal boundaries during mapping. Extensive experiments on public and custom datasets show that PI-EAT-SLAM achieves superior tracking accuracy and highly competitive novel-view synthesis compared to state-of-the-art baselines.
Demo video coming soon...
Physics-informed thermal preprocessing that maps raw 14/16-bit intensities to radiance via the Stefan-Boltzmann law (L ∝ T⁴). This non-linear rescaling boosts true thermal contrast while suppressing noise — a far cleaner foundation for edge extraction than linear min-max scaling.
Gradient-aware Edge-KLT tracking combines KLT with high-confidence Canny edges. A three-stage pipeline — distance filtering, dual-constraint (spatial + gradient-directional) matching, and residual rejection — keeps only reliable correspondences under severe thermal noise.
A novel edge-aware smoothness loss regularizes mapping: it strongly smooths flat regions while strictly preserving sharp thermal discontinuities, enforcing thermal equilibrium for high-fidelity dense reconstruction.