Curiosity-Driven Attention for Anomaly Road Obstacles Segmentation in Autonomous Driving

Abstract

The inability of semantic segmentation methods to detect anomaly road obstacles not pre-defined in the datasets significantly hinders the safety-critical application in autonomous driving. The excessively complex anomaly detection approaches cannot accommodate the constraints on the inference time of intelligent vehicles. Inspired by the fact that humans have a natural instinct to be curious about unknown objects in a new environment, we propose a novel curiosity-driven attention mechanism (CuDAM) for anomaly road obstacles segmentation. CuDAM adopts the attention map as a new uncertainty judging criterion and utilizes it to improve the efficiency of the model. Specifically, CuDAM is composed of three parts: 1) an attention module for generating an attention map; 2) a reward mechanism for encouraging the network to focus its attention on uncertain regions; 3) an attention loss function for widening the distance between the attention values of deterministic and uncertain pixels. Different from previous approaches, CuDAM can improve both anomaly detection and semantic segmentation performance without complex operations and training, which makes it widely applicable to existing semantic segmentation models. The result of qualitative and quantitative experiments shows that such a straightforward approach achieves consistent significant improvements in anomaly detection performances with the various uncertainty estimation methods, demonstrating the broad applicability of CuDAM.