Small Anomaly Segmentation in Autonomous Driving

Abstract

Identifying unfamiliar or unusual objects on the road poses a significant challenge in autonomous driving. While recent studies have achieved high accuracy in identifying anomalies of regular size, the detection of smaller objects remains a more complex problem. Here, we introduce AutoFocusAnomaly (AFA), a practical approach designed to enhance the detection of small anomalies. AFA integrates a modified version of the AutofocusFormer segmentation model with the classic uncertainty estimation function, particularly the maximum logit (i.e., the highest values among classes in the model's output). To assess the performance of the method, we take a portion of the Lost And Found (LAF) dataset to render it suitable as a new dataset called LAF Far (LAFF) for small anomaly segmentation. Results show the effectiveness of our method in anomaly segmentation. Specifically in the small anomaly segmentation task, we obtain the highest Average Precision (AP) coupled with a competitively low false positive rate, which is significantly better than State-Of-The-Art(SOTA) methods. We believe that these might shed light on future research in the domain of small anomalies segmentation.