VT-Intrinsic Dataset

A collection of aligned visible–thermal image pairs of real-world scenes captured across diverse illumination conditions and surface textures, released with the paper:

VT-Intrinsic: Physics-Based Decomposition of Reflectance and Shading using a Single Visible-Thermal Image Pair (CVPR 2026, Carnegie Mellon University)

• Project page: https://vt-intrinsic.github.io
• Paper (CVF Open Access): https://openaccess.thecvf.com/content/CVPR2026/papers/Yuan_VT-Intrinsic_Physics-Based_Decomposition_of_Reflectance_and_Shading_using_a_Single_CVPR_2026_paper.pdf

Overview

The VT-Intrinsic dataset provides 623 spatially aligned visible–thermal image pairs of stationary outdoor scenes including parks, schools, cathedrals, plazas, museums, and urban streets. It targets intrinsic image decomposition (IID) — separating a visible image into reflectance (albedo) and shading — using a thermal image as a physically grounded auxiliary cue.

The motivation: light not reflected by an opaque surface is absorbed and re-emitted as heat, observable by a long-wave infrared (LWIR) camera. The visible–thermal intensity ordinality at and between any pair of pixels therefore constrains the underlying albedo/shading ordinality without requiring ground-truth labels (see paper, Sec. 3).

Compared to existing IID datasets (which are predominantly indoor or synthetic and lack a thermal modality) and existing visible–thermal datasets (which target detection/tracking on vehicles or people rather than light–heat physics), VT-Intrinsic contributes:

1. A large collection of high-quality, real-world outdoor visible–thermal pairs with diverse albedo–shading combinations.
2. Scenes with dense pseudo-ground-truth ordinalities usable as supervision for learning-based intrinsic image decomposition.

Capture Setup

Component	Spec
Thermal camera	FLIR Boson, 512×640, 24° HFOV, ≤50 mK NEDT, LWIR (8–14 μm)
Visible camera	IDS UI-3130, 600×800, 27° HFOV
Co-location optics	Gold dichroic mirror (BSP-DI-25-2)
For some distant scenes	Side-by-side mounted cameras with per-scene homography

Visible processing. For each scene we captured 20 exposure-bracketed RAW frames, applied edge-aware demosaicing, and merged into a linear HDR image (Debevec & Malik '97). Five HDR frames were averaged to suppress sensor noise.

Cross-modal alignment. The HDR visible image and the thermal image were further aligned via a per-scene homography.

Contents

VT-Intrinsic_dataset/
├── vis_thr_pairs/      # raw linear captures
│   ├── DDMM_HHMMSS_linear_vis.npy
│   └── DDMM_HHMMSS_thr.npy
└── visualization/         # tonemapped and colormapped previews
    ├── DDMM_HHMMSS_vis.png
    └── DDMM_HHMMSS_thr.png

Folder	Files	Format	Description
vis_thr_pairs/	1246 (623 pairs)	.npy	Linear visible image (3-channel float) and aligned thermal image (1-channel float).
visualization/	1246 (623 pairs)	.png	Gamma-tonemapped visible and false-color thermal renderings, for quick visual inspection.

Intended Use

• Physics-based or learning-based (e.g., w/ generative prior) intrinsic image decomposition (IID) with a thermal auxiliary channel.
• Pseudo-ground-truth supervision for IID models or broader reflectance and lighting understanding (see paper Sec. 4 for the local edge and non-local point-pair ordinalities).
• Research on multimodal scene understanding on real outdoor scenes.
• And more ...

Limitations

Uncooled microbolometer thermal cameras offer limited SNR, especially in low-light or fast-moving scenes.

Citation

If you find this dataset helpful, please cite:

@inproceedings{yuan2026vtintrinsic,
  title     = {{VT-Intrinsic}: Physics-Based Decomposition of Reflectance and Shading using a Single Visible-Thermal Image Pair},
  author    = {Yuan, Zeqing and Ramanagopal, Mani and Sankaranarayanan, Aswin C. and Narasimhan, Srinivasa G.},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2026}
}

License

Released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license unless noted otherwise. You are free to share and adapt the data for any purpose with attribution.