Augmentation-Based Approaches for Overcoming Low Visibility in Street Object Detection

Road object detection in low-visibility conditions, such as nighttime, fog, and rain, is difficult for standard machine learning models, which often struggle because of limited training data. The collection of comprehensive datasets that encompass all possible scenarios encountered during deployment is often impractical in terms of time and cost. To overcome this limitation, this work proposes the use of specific augmentations tailored to address the challenges associated with low-visibility conditions. The model employed in this research was trained on sub-sets that complemented the missing low-visibility circumstances. Augmentations based on depth and Fourier domain techniques were applied to simulate such conditions during training and enhance the model's performance when faced with such a scenario. Experimental results demonstrated that appropriately applied augmentations can improve the model's performance. Specifically, in rainy weather, the best-performing model trained on augmented data achieved a 3.4% improvement over a model trained on non-augmented data. In other cases, however, the proposed augmentations did not increase significantly the per-formance of the classifier.