ERUP-YOLO: Enhancing Object Detection Robustness for Adverse Weather Condition by Unified Image-Adaptive Processing

TL;DR

ERUP-YOLO addresses the challenge of object detection under adverse weather by introducing two unified, differentiable image-processing filters, BPW and KBL, assembled via an image-processing parameter encoder. The BPW filter maps pixel intensities through a cubic Bézier curve, enabling a flexible, differentiable transformer of global image tone, while the KBL filter uses locally linear convolutions to capture defog and sharpen effects with shared per-channel parameters. A domain-agnostic BPW-based data augmentation further enhances robustness without condition-specific customization. Empirically, ERUP-YOLO achieves state-of-the-art performance on foggy and low-light datasets, outperforming prior image-adaptive methods across VOC, RTTS, ExDark, and DAWN; ablations show that the joint BPW+KBL configuration provides the strongest gains while highlighting scenarios where global brightness can be a limitation. Overall, the approach offers a compact, customization-free preprocessing front-end that improves real-world detection robustness under diverse adverse conditions.

Abstract

We propose an image-adaptive object detection method for adverse weather conditions such as fog and low-light. Our framework employs differentiable preprocessing filters to perform image enhancement suitable for later-stage object detections. Our framework introduces two differentiable filters: a Bézier curve-based pixel-wise (BPW) filter and a kernel-based local (KBL) filter. These filters unify the functions of classical image processing filters and improve performance of object detection. We also propose a domain-agnostic data augmentation strategy using the BPW filter. Our method does not require data-specific customization of the filter combinations, parameter ranges, and data augmentation. We evaluate our proposed approach, called Enhanced Robustness by Unified Image Processing (ERUP)-YOLO, by applying it to the YOLOv3 detector. Experiments on adverse weather datasets demonstrate that our proposed filters match or exceed the expressiveness of conventional methods and our ERUP-YOLO achieved superior performance in a wide range of adverse weather conditions, including fog and low-light conditions.

Figures (11)

• Figure 1: Failure cases of the conventional YOLOv3 detector on adverse weather images (left side) such as fog, rain, low-light and snow conditions, and the successful detection results after applying our proposed image processing filters (right side).
• Figure 2: (a) Overview of the proposed ERUP-YOLO framework, which integrates the Bézier curve-based pixel-wise (BPW) filter and the kernel-based local (KBL) filter before the object detection network. (b) Illustration of how the proposed BPW and KBL filters unify and generalize the conventional image processing filters.
• Figure 3: Plots illustrating the input-output pixel intensity mappings of conventional pixel-wise filters (gamma, contrast, tone) compared to the proposed Bézier curve-based pixel-wise (BPW) filter.
• Figure 4: Parameters definition of BPW filter.
• Figure 5: Comparison of prior data-specific augmentation (2nd column) and the proposed Bézier curve-based pixel-wise (BPW) augmentation (3rd column) for low-light (top row) and foggy (bottom row) conditions. The 3rd column shows BPW augmentation manually adjusted to mimic the prior augmentations. The rightmost column visualizes the pixel intensity mappings as plots.