Uncertainty Estimation in Deep Neural Object Detectors for Autonomous Driving

Abstract: One of the most challenging problems in autonomous driving, and in computer vision in general, is the task of object detection. Recently, the advances of deep learning in the domain of computer vision have led the research community to apply ideas from deep learning to the problem of object detection. Such frameworks have shown promising results by outperforming all of the traditional approaches based on hand-crafted features. However, deep neural networks fail to provide reliability measurement for their predictions, which is a crucial requirement for critical systems. Moreover, it has been shown that deep neural networks tend to give more overconfident predictions compared to their shallow counterpart. Our objective with the current degree project is to review popular deep neural object detection frameworks and to investigate methods for uncertainty estimation in deep convolutional neural networks. Additionally, we would like to find out if applying a method of uncertainty estimation could improve the performance of a deep neural object detector in terms of standard object detection metrics such as average precision and log average miss rate. Moreover, we aim at discovering if an approach for measuring reliability improves the calibration quality of an object detector. Finally, we focus on investigating if uncertainty estimates could be utilised for detecting misclassified examples or highly occluded/truncated objects within an image. In order to answer these questions, we construct a variety of experiments that investigate different setups and conditions. Overall, we found that uncertainty estimation is a promising approach for alleviating overconfidence in deep neural networks as well as for providing a reliability measurement for all predictions.