From Fog to Failure: The Unintended Consequences of Dehazing on Object Detection in Clear Images

TL;DR

The study investigates the unintended consequences of integrating dehazing into object detection by proposing a ROI-guided, spatial-attention dehazing pipeline (AOD-NetX) that feeds a heavier detector. While effective in foggy conditions, the approach unexpectedly degrades performance on clear images due to feature-space misalignment and domain adaptation bias. The method combines a lightweight preliminary detector, region-focused dehazing using the transmission map $K(x)$ refined to $K'(x)$, and a final robust detector, with evaluations on Foggy Cityscapes and RESIDE-$\beta$ showing mixed results across in-distribution and out-of-distribution data. The findings highlight the need for selective preprocessing, context-aware enhancement, and joint optimization to balance enhancement gains against potential losses in clear conditions. Practical implications include guiding the design of hybrid perception pipelines that avoid universal benefits from cascading transformations and consider real-time feasibility.

Abstract

This study explores the challenges of integrating human visual cue-based dehazing into object detection, given the selective nature of human perception. While human vision adapts dynamically to environmental conditions, computational dehazing does not always enhance detection uniformly. We propose a multi-stage framework where a lightweight detector identifies regions of interest (RoIs), which are then improved via spatial attention-based dehazing before final detection by a heavier model. Though effective in foggy conditions, this approach unexpectedly degrades the performance on clear images. We analyze this phenomenon, investigate possible causes, and offer insights for designing hybrid pipelines that balance enhancement and detection. Our findings highlight the need for selective preprocessing and challenge assumptions about universal benefits from cascading transformations.

Figures (6)

• Figure 1: Overall architecture of Perceptual Piercing: (a) Preliminary detection using lightweight object detection model (b) Gaze-directed dehazing using spatial attention on region of interests (c) Final detection using a large and robust model.
• Figure 2: Architecture of AOD-NetX: The model takes the transmission map output, $K(x)$, from AOD-Net and applies a spatial attention layer to emphasize key regions of interest (bounding boxes) in the input image. The refined transmission map, $K'(x)$, is then utilized to dehaze the image.
• Figure 3: Comparison of mean Average Precision (mAP) for different dehazing and object detection module combinations.
• Figure 4: Dehazing performance on Foggy Cityscapes dataset.
• Figure 5: Dehazing performance on Foggy Cityscapes dataset.