Pixel-Wise Detection of Road Obstacles on Drivable Areas by Introducing Maximized Entropy to Synboost Framework

Abstract

Semantic segmentation plays a crucial role in understanding the surroundings of a vehicle in the context of autonomous driving. Nevertheless, segmentation networks are typically trained on a closed-set of inliers, leading to misclassification of anomalies as in-distribution objects. This is especially dangerous for obstacles on roads, such as stones, that usually are small and blend well with the background. Numerous frameworks have been proposed to detect out-of-distribution objects in driving scenes. Some of these frameworks use softmax cross-entropy measurements as an attention mechanism for a dissimilarity network to find anomalies. However, a significant limitation arises from the segmentation network’s tendency toward overconfidence in its predictions, resulting in low cross-entropy in regions where anomalies are present. This suggests that normal cross-entropy is a low-quality prior for anomaly detection. Therefore, for the task of detecting stones on roads, we propose utilizing a fined-tuned segmentation network with a changed target, from semantic segmentation to maximize the cross-entropy in anomalous areas. With this, we feed the dissimilarity network with a better prior image. Furthermore, due to the lack of datasets with enough samples of stones for pixel-wise detection, we synthetically added stones on images of driving scenes to create a dataset for fine-tuning and training. The results of our comparative experiments showed that our model attains the highest average precision while having the lowest false positive rate at 95% true positive rate when evaluating on a real-stone image dataset.