Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops (ICCVw 2025) · 2025

Vision At Night - Exploring Biologically Inspired Preprocessing For Improved Robustness Via Color And Contrast Transformations

Lorena Stracke, Lia Nimmermann, Shashank Agnihotri, Margret Keuper, Volker Blanz

Why this publication matters

Darkness changes the information available to a camera, while many vision models are trained mainly on ordinary color photographs. This work explores simple transformations inspired by how biological vision handles color and contrast. It shows that changing the representation of the input can improve robustness without making the learned model more complicated.

Abstract

Deep neural networks surpass humans on many vision benchmarks, yet remain far less robust to distribution shifts such as illumination and weather changes. Existing approaches address this challenge by additional training data, extensive augmentation, architectural modifications, or test-time adaptation. In this work, we explore a complementary direction: inspired by the human retina, we propose a fixed, model-agnostic preprocessing module that extracts signals that are more stable with respect to variations of illumination. Our method combines color remapping with local contrast extraction, producing sparse representations that emphasize structural features. We study its impact on semantic segmentation by training on Cityscapes and evaluating generalization under adverse conditions on Dark Zurich and ACDC. Our results show that the biologically inspired preprocessing preserves in-distribution performance while consistently improving robustness in challenging lighting scenarios, such as nighttime, where annotated training data are scarce. Moreover, the segmentation accuracy remains stable even when the contrast-based representation is sparsified by up to 70%. These gains suggest that rethinking the input representation itself can improve robustness while also opening opportunities for lower-latency, transmission-aware imaging sensors when sparsity can be exploited close to acquisition.

Abstract source ↗

Figures

Nighttime segmentation using alternative color and contrast representations. — **Figure 1.** Robustness can begin before the neural network. In-distribution (above, Cityscapes [9]) versus OOD low-light (below, ACDC Night [44]) results for UPerNet. The leftmost column shows the baseline, which transfers poorly to ACDC Night, while our biologically inspired preprocessing in all considered variants (luminance, color-opponency, and single-color) significantly improves results across all labels by preserving robust structural cues. View in source ↗

**Figure 2.** Overview of the four evaluated pipelines. In comparison with a baseline (top row), we consider our preprocessing in three variants: luminance, color-opponency and single-color. Our preprocessing begins with a color remapping (1) from RGB values to three different color spaces. Then follows the contrast extraction (2), where we simulate a convolution with a DoG-like kernel. Finally, we enforce sparsity (3) by setting low absolute values to zero, highlighted in pink. Each variant produces 3 channels that serve as a new input to different neural network architecures (right). View in source ↗

Cite this paper

Download .bib

@inproceedings{stracke2025visionatnight72,
  title = {{Vision At Night - Exploring Biologically Inspired Preprocessing For Improved Robustness Via Color And Contrast Transformations}},
  author = {Lorena Stracke and Lia Nimmermann and Shashank Agnihotri and Margret Keuper and Volker Blanz},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops (ICCVw 2025)},
  year = {2025},
  url = {https://arxiv.org/pdf/2509.24863}
}

Figures and abstract are reproduced from the linked research sources. Credit remains with the authors and publishers.