Official repository for the paper accepted at the 19th International Conference on Control, Automation, Robotics and Vision (ICARCV 2026).
Alexandre Riffard1 · Mathieu Labussiere1 · Pierre Duthon2 · Romuald Aufrere1
1 Université Clermont Auvergne, Clermont Auvergne INP, CNRS, Institut Pascal, Clermont-Ferrand, France
2 Cerema, Research Team "Intelligent Transport Systems", Clermont-Ferrand, France
✉️ Contact: {alexandre.riffard, mathieu.labussiere, romuald.aufrere}@uca.fr, pierre.duthon@cerema.fr
Note
This repository is a placeholder accompanying our ICARCV 2026 paper.
The full source code, pre-trained models, and instructions will be made publicly available upon official publication of the paper.
Reliable perception in adverse weather, such as dense fog, heavy rain, or severe glare, remains a critical challenge for Autonomous Vehicles (AV).
While visible (VIS) cameras degrade rapidly in these conditions, Short-Wave Infrared (SWIR) sensors offer enhanced atmospheric penetration and structural retention.
However, fusing VIS and SWIR is hampered by sensor noise, modality misalignment, and the lack of pixel-perfect ground truth.
In this paper, we propose Diff-VISWIR, a prior-guided, end-to-end differentiable fusion architecture.
Moving away from purely data-driven architectures, our method unrolls a physical Laplacian pyramid into a learnable network.
It introduces a modality-specific saliency extractor and a soft-veto mechanism that dynamically suppresses the VIS contribution when robust SWIR structures are detected.
The architecture is explicitly constrained by physical and colourimetric bounds, and optimised via a semantic-aware multi-objective loss that includes supervision from a downstream object detector.
Experimental results demonstrate that Diff-VISWIR achieves robust zero-shot domain generalisation.
On a severely degraded dataset, it delivers top-tier downstream detection performance, achieving high precision while mitigating the formation of false-positive contours under adverse conditions.
Furthermore, our framework operates at 33-37 frames per second on high-definition automotive inputs, satisfying real-time deployment constraints.
If you find this work useful in your research, please cite our paper:
This work was supported by the International Research Center "Innovation Transportation and Production Systems" of the I-SITE CAP 20-25.
We thank the Institut Pascal (Université Clermont Auvergne, Clermont Auvergne INP, CNRS) and Cerema (Team Intelligent Transport Systems) for their support.
This project is licensed under the GNU Lesser General Public License v3.0 (LGPL-3.0) - see the LICENSE file for details.