Abstract
This paper introduces DB-GAN (Detector-Boost GAN), a novel framework designed to normalize image illumination specifically for machine vision tasks. Unlike traditional image enhancement methods that prioritize human aesthetics, DB-GAN integrates a frozen object detector into its training loop to minimize detection loss. This approach bridges the domain gap caused by extreme lighting—such as harsh shadows or overexposure—significantly boosting the performance of standard detectors in challenging real-world environments without requiring extensive re-annotation of datasets.
Introduction
Object recognition has achieved remarkable accuracy thanks to deep learning, yet it remains notoriously fragile when faced with real-world lighting variations. Extreme conditions, such as the blinding glare of direct sunlight or the grainy darkness of poorly lit rooms, create a significant "domain gap" that standard detectors struggle to bridge. Existing solutions typically involve either classical contrast normalization, which fails to generalize to complex scenes, or massive data collection, which is prohibitively expensive for tasks like 6D pose estimation. DB-GAN addresses this by automatically "fixing" the lighting in an image before it reaches the detector, acting as a smart pre-processor that makes the vision pipeline robust to environmental changes.
Methods: Detection-Driven Reconstruction
The core of the framework is an architecture designed for "Detection-Driven Reconstruction." The system utilizes an encoder-decoder generator (similar to a U-Net) that takes an image with poor lighting and outputs a version with uniform, normalized illumination.
The innovation lies in the multi-part loss function. Beyond standard reconstruction and adversarial losses, DB-GAN incorporates a Detection Loss. During training, the generator's output is fed into a pre-trained object detector (such as SSD or SSD6D) with frozen weights. The error from the detector is back-propagated into the generator, forcing it to prioritize fixing specific lighting artifacts that cause detection failure. This creates a specialized "translator" between harsh real-world environments and the idealized uniform lighting the detector was originally trained on.
Experimental Evaluation
The authors evaluated DB-GAN on several challenging benchmarks, including the Toyota Light and TUD Light datasets. They also introduced the Toyota TrueBlue dataset to specifically test robustness against white balance and color temperature changes (ranging from 2500K to 10000K).
Quantitative results showed substantial improvements across both 2D detection and 6D pose estimation:
- 2D Detection: On the Toyota Light dataset, the baseline SSD detector's mean Average Precision (mAP) jumped from 0.27 to 0.72 with DB-GAN normalization.
- TUD Light Performance: DB-GAN improved mAP from 0.18 to 0.66, outperforming state-of-the-art enhancement methods like EnlightenGAN and Retinex-Net.
- Color Robustness: On the Toyota TrueBlue dataset, DB-GAN achieved 0.73 mAP, nearly doubling the baseline performance of 0.39.
Conclusion
DB-GAN demonstrates the power of machine-centric normalization. By treating a detector's failure as a training signal for image reconstruction, the system achieves remarkable zero-shot generalization from synthetic training data to real-world scenarios. While the GAN pre-processor adds some computational overhead, the significant gains in robustness make it a compelling approach for autonomous systems operating in unpredictable lighting. Future work may expand this "detector-boosted" approach to handle other environmental variations like fog, rain, or motion blur.
Summary updated