Road Obstacle Detection based on Unknown Objectness Scores

TL;DR

A novel anomaly score is proposed by integrating the object-detection fashions into the pixel-wise anomaly detection methods by incorporating the object-detection fashions into the pixel-wise anomaly detection methods to achieve stable performance for detecting unknown objects.

Abstract

The detection of unknown traffic obstacles is vital to ensure safe autonomous driving. The standard object-detection methods cannot identify unknown objects that are not included under predefined categories. This is because object-detection methods are trained to assign a background label to pixels corresponding to the presence of unknown objects. To address this problem, the pixel-wise anomaly-detection approach has attracted increased research attention. Anomaly-detection techniques, such as uncertainty estimation and perceptual difference from reconstructed images, make it possible to identify pixels of unknown objects as out-of-distribution (OoD) samples. However, when applied to images with many unknowns and complex components, such as driving scenes, these methods often exhibit unstable performance. The purpose of this study is to achieve stable performance for detecting unknown objects by incorporating the object-detection fashions into the pixel-wise anomaly detection methods. To achieve this goal, we adopt a semantic-segmentation network with a sigmoid head that simultaneously provides pixel-wise anomaly scores and objectness scores. Our experimental results show that the objectness scores play an important role in improving the detection performance. Based on these results, we propose a novel anomaly score by integrating these two scores, which we term as unknown objectness score. Quantitative evaluations show that the proposed method outperforms state-of-the-art methods when applied to the publicly available datasets.

Figures (5)

• Figure 1: Examples of road obstacle detection results using the proposed and representative methods. The proposed anomaly score---unknown objectness score---reduces the number of false-positive predictions, especially in the background regions, compared to the baseline scores (c) and the results of the existing methods (d,e).
• Figure 2: Overview of the proposed method. The input image is fed to a semantic-segmentation network with a sigmoid head. Unknown objectness scores are obtained from the sigmoid outputs by combining unknown and objectness scores.
• Figure 3: Venn diagrams comparing between the unknown score (Eg. \ref{['eq:one_vs_rest_score']}) and the unknown objectness score (Eg. \ref{['eq:anomaly_score']}). Typical objects and backgrounds indicate predefined semantic-segmentation classes, while unknown objects and backgrounds indicate ones that are not fall into the predefined classes. Road obstacles are classified into unknown objects.
• Figure 4: Examples of labels for three typical types of pixels. This figure shows three pixels $i,j,k$ corresponding to a car, a road and a road obstacle pixels, respectively. Pixel $i$ is assigned to both the vehicle and object classes, while pixel $j$ is assigned to the road class only, because cars are objects, and roads are not. Pixel $k$ is assigned to the object class only, because road obstacles do not fall into any of the predefined classes, while they are objects.
• Figure 5: Qualitative results for Ourtlier Exposure, Synboost, and the unknown objectness score (UOS).