Title: TeX-1500: A Paired Real-World LWIR Hyperspectral Dataset and Benchmark for Temperature-Emissivity-Texture Decomposition

Abstract:

The process of Temperature-Emissivity-Texture (TeX) decomposition aims to reconstruct an object’s thermal state, material spectral signature, and visible-like geometric details using Long-Wave Infrared Hyperspectral Imaging (LWIR HSI). However, current TeX methodologies predominantly rely on scene-specific inverse solvers. This approach has been hindered by the absence of paired LWIR HSI-TeX supervision, which has restricted the development of learning-based decomposition techniques. To bridge this critical gap, we present TeX-1500, a comprehensive benchmark and large-scale dataset featuring paired LWIR HSI-TeX samples designed for supervised HSI-to-TeX decomposition.

TeX-1500 comprises 1,522 calibrated pairs derived from real-world scenes. These datasets are sourced from DARPA Invisible Headlights (DARPA IH) pushbroom imagery alongside our own Fourier Transform Infrared Spectroscopy (FTIR) acquisitions. The collection spans five distinct locations, encompasses all four seasons, and includes varied acquisition times, heterogeneous wavelength configurations, and two different sensor families. Every sample within the dataset contains a calibrated radiance cube restricted to valid bands, precise calibrated wavelength positions, and aligned supervision labels for temperature, emissivity, and texture. These labels are generated via a unified protocol for restoration and TeX construction.

Additionally, we introduce TeX-UNet, a straightforward baseline model that is wavelength-aware. This network maps calibrated HSI bands and their corresponding wavelength positions to TeX fields. Our experimental results, conducted on held-out DARPA IH pushbroom scenes and through zero-/few-shot transfer to FTIR scenes, demonstrate that TeX-1500 offers effective paired supervision. It establishes a quantifiable benchmark for data-driven thermal perception centered on physical properties.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC