International Conference on Computer and Knowledge Engineering

Home / 13th International Conference on Computer and Knowledge Engineering

Standardized ReACT Logits: An Effective Approach for Anomaly Segmentation in Self-driving Cars

Authors :

Mahdi Farhadi¹ Seyede Mahya Hazavei² Shahriar Baradaran Shokouhi³

1- School of Electrical Engineering Iran University of Science and Technology Tehran, Iran 2- School of Electrical Engineering Iran University of Science and Technology Tehran, Iran 3- School of Electrical Engineering Iran University of Science and Technology Tehran, Iran

Keywords :

anomaly segmentation،autonomous driving،semantic segmentation

Abstract :

The identification of unexpected road objects is a crucial aspect in the field of autonomous driving. Various methods have been proposed for anomaly segmentation, which can be categorized into three important categories: the use of auxiliary datasets, the utilization of uncertainty maps, and the reconstruction networks. In this study, the DeepLabv3+ network serves as the primary semantic segmentation model. We calculate the energy function, a type of uncertainty, before the output of the upsampling layer and employ it as an anomaly score. Unlike samples within the distribution, samples outside the distribution do not exhibit deviations from the standard criteria in terms of the distribution of these values. To address this issue, we use the ReAcT operator, which replaces values exceeding a threshold with the threshold value, leading to improved performance. The proposed method enhances the performance of anomaly segmentation by incorporating a single step of standardizing anomaly scores and considering the semantic dependencies of pixels in each region. This is achieved through two steps: Iterative Boundary Suppression and Dilated Smoothing. Evaluation on three common datasets in the field, namely Fishyscapes Lost & Found, Fishyscapes Static, and Road Anomaly, demonstrates the method's robustness in anomaly segmentation without the need for auxiliary datasets or network retraining.