PI-EAT-SLAM: Physics-Informed Edge-Aware Thermal Gaussian Splatting SLAM (IROS 2026)

Abstract

Edge-driven SLAM for the thermal domain

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.

🌡️

Physics-informed contrast

Stefan-Boltzmann radiance rescaling (L ∝ T⁴) restores true thermal contrast while suppressing sensor noise.

⚡

Robust edge tracking

GE-KLT couples KLT tracking with nearest-neighbor matching under spatial + gradient-directional dual-constraint outlier rejection.

🧊

Edge-aware mapping

An edge-aware smoothness loss enforces thermal equilibrium while preserving sharp boundaries in dense 3DGS.

Method

Three pillars, one robust system

Stefan-Boltzmann Thermal Rescaling

Raw 14/16-bit thermal data must be compressed to 8 bits for tracking. Conventional min-max scaling loses critical thermal detail and contrast. We instead convert raw intensities to an absolute temperature map and apply the Stefan-Boltzmann law to obtain surface radiance, which is proportional to the fourth power of temperature. This non-linear mapping amplifies subtle but physically meaningful temperature variations far better than linear scaling — yielding sharper, less noisy 8-bit images. CLAHE and bilateral filtering then sharpen local detail while attenuating noise.

L(x, y) = ε · σ · T(x, y)⁴

Edge-point matching comparison across rescaling methods

Why physics matters for edges. Edge-point matching across rescaling methods: (a) original thermal, (b) min-max scaling, (c) FieldScale, and (d) our Stefan-Boltzmann rescaling. Our method produces the most reliable, consistent matches (red/blue dots with cyan residual lines) by enhancing contrast while suppressing noise.

Gradient-aware Edge-KLT (GE-KLT) Tracking

Pure photometric alignment is fragile in thermal imagery. Our multi-stage matching pipeline combines KLT tracking (local patch appearance) with high-confidence Canny edges through three sequential outlier-rejection stages: (1) distance-based filtering, (2) dual-constraint matching that requires both spatial proximity and gradient-directional agreement, and (3) residual-based rejection. This strictly anchors matches to true physical boundaries, dramatically improving pose-estimation robustness in contrast-limited scenes.

GE-KLT residual. A rendered edge point (red) is KLT-tracked to a candidate (blue), then matched to the nearest valid ground-truth Canny edge (green) under dual spatial + gradient constraints. The residual (magenta) couples spatial distance with gradient angular difference — anchoring every match to a real edge.

Edge-aware Smoothness for Mapping

Standard smoothness losses oversmooth important geometric boundaries. Our edge-aware smoothness loss applies strong smoothing in flat regions while strictly preserving sharp thermal discontinuities, enforcing natural thermal equilibrium. A sigmoid-weighted edge term down-weights regularization exactly where strong radiance gradients occur, so the final 3DGS map stays crisp at object boundaries and clean in homogeneous regions.

L_e-smooth = (1 / 4M) · Σ_i,j ( |R_i+1,j − R_i,j| + |R_i,j+1 − R_i,j| ) · (1 − w_edge(i, j))

w_edge(i, j) = σ( α(|∇R_i,j| − β) ), σ(x) = 1 / (1 + e^−x), α = 8, β = 0.1

Here R_i,j is the rendered thermal radiance at pixel (i, j) after rescaling and M is the total pixel count; the sigmoid edge weight w_edge ≈ 0 in flat regions (strong smoothing) and ≈ 1 near strong edges (boundaries preserved). The final mapping loss aggregates the photometric loss, our edge-aware smoothness loss, and isotropic regularization over keyframes k in the window W:

L_mapping = Σ_k∈W ( λ_pho L^k_pho + λ_e-smooth L^k_e-smooth ) + λ_iso L_iso

Evaluation

Superior tracking, competitive rendering

Trajectory comparison on the multi-spectral dataset

Trajectory accuracy. Estimated trajectories on the multi-spectral dataset — Ours vs. MonoGS, compared against Ground Truth (black). Red boxes highlight segments where MonoGS drifts significantly, whereas PI-EAT-SLAM maintains accurate alignment.

Edge-matching quality. GE-KLT (left) vs. plain E-KLT (right). By explicitly anchoring matches to valid ground-truth edge points (orange boxes), GE-KLT suppresses erroneous matches and prevents tracked points from drifting away from real physical boundaries — yielding highly consistent residual vectors.

Camera Tracking Accuracy — ATE RMSE (m) on the Multi-spectral Dataset. **Bold** = best, underline = second best.
Sequence	ORB2 (RGB-D)	ORB2 (Thermal-D)	MonoGS (Thermal-D)	Thermal-D Odom	Ours
Bright
ds1-xyz	0.0119	0.0832	0.0641	0.0719	0.0541
ds1-xyz-ps	0.0402	0.1412	0.0419	0.0154	0.0314
ds1-hfsp	0.1061	0.4416	0.0959	0.0582	0.0485
ds2-xyz	0.0618	Fail	Fail	0.0375	0.0865
ds2-hfsp	0.0759	0.1525	0.1536	0.0288	0.1435
Varying (illumination changes)
ds1-xyz-ic	Fail	0.2502	0.0815	0.1151	0.0553
ds1-xyz-ic-ps	Fail	0.3490	0.0609	0.0778	0.0520
ds1-hfsp-ic	Fail	0.4674	0.0797	0.0514	0.0452
ds2-xyz-ic	Fail	0.2032	0.1705	0.1136	0.0840
ds2-hfsp-ic	Fail	0.1889	0.0988	0.1099	0.0626

Lower is better. RGB-D ORB-SLAM2 fails on every Varying sequence, while PI-EAT-SLAM (Ours) leads across all of them and stays competitive under Bright conditions.

Novel-view synthesis. From left to right: Ground Truth, MonoGS, and Ours. MonoGS blurs object boundaries (red boxes) and produces ghosting/black artifacts in flat regions (orange boxes), whereas our edge-aware mapping preserves sharp structures and renders clean, artifact-free surfaces.

Novel View Synthesis — Custom Dataset (Monocular). **Bold** = best, underline = second best.
Method	Metric	Easy			Difficult
Method	Metric	xyz	sfm	circle	upper_turn	down_turn	up_down
MonoGS	PSNR ↑	25.54	27.22	21.81	15.31	11.36	12.66
	SSIM ↑	0.941	0.943	0.876	0.528	0.268	0.284
	LPIPS ↓	0.179	0.145	0.207	0.540	0.690	0.656
HI-SLAM2	PSNR ↑	23.66	24.41	22.36	19.43	18.54	20.10
	SSIM ↑	0.916	0.895	0.885	0.855	0.846	0.865
	LPIPS ↓	0.193	0.196	0.207	0.464	0.467	0.400
Ours (Mono)	PSNR ↑	25.42	27.29	23.73	20.82	11.85	15.26
	SSIM ↑	0.944	0.944	0.920	0.870	0.326	0.436
	LPIPS ↓	0.168	0.141	0.158	0.378	0.661	0.574

PSNR/SSIM higher is better, LPIPS lower is better. Our preprocessing intentionally alters intensity distributions to prioritize tracking, yet PI-EAT-SLAM still tops NVS quality on most Easy sequences and the upper_turn case, remaining competitive on the hardest rotational trajectories.

PI-EAT-SLAM