Reducing Overconfidence Predictions in Autonomous Driving Perception

Abstract

Object recognition is a crucial step in perception systems for autonomous and intelligent vehicles, as evidenced by the numerous research works in the topic. In this paper, object recognition is explored by using multisensory and multimodality approaches, with the intention of reducing the false positive rate (FPR). The reduction of the FPR becomes increasingly important in perception systems since the misclassification of an object can, depending on the circumstances, potentially cause accidents. In particular, this work presents a strategy through Bayesian inference to reduce the FPR considering the likelihood function as a cumulative distribution function from Gaussian kernel density estimations, and the prior probabilities as cumulative functions of normalized histograms. The validation of the proposed methodology is performed on the KITTI dataset using deep networks (DenseNet, NasNet, and EfficientNet), and recent 3D point cloud networks (PointNet, and PointNet++), by considering three object-categories (cars, cyclists, pedestrians) and the RGB and LiDAR sensor modalities.

Introspection of DNN-Based Perception Functions in Automated Driving Systems: State-of-the-Art and Open Research Challenges

“…Then such methods can query to introspect the decisions of the classification model. Confidence Calibration [37], [48], [49], [50], [51], [52], [53], [54] Use of confidence outputs, or logits from neural network to process them further.…”

“…Use of confidence outputs, or logits from neural network to process them further. Maximum likelihood and maximum-a-posteriori are used in [54].…”

“…They used them to improve an object detector, SSD [87], by eliminating unreliable predictions in open-set conditions, where the operating environment is not restricted to the categories presented in the training set. In [54], utilising maximum likelihood and maximum-a-posteriori functions rather than softmax and sigmoid is introduced. They evaluated the effect of the functions on the confidence representation of object detection.…”

“…Confidence Estimation [54], [90], [91], [92], [93], [94] A CNN is utilised to map input to feature space for estimating von Mises-Fisher Distribution. In testing, the likelihood is used as a novelty score.…”

“…They evaluated their proposed solution with both adversarial and out-of-distribution samples to show its efficiency in different problematic cases. Alternatively, in [54], the softmax function is replaced with maximum likelihood and maximum-a-posteriori functions for better probabilistic confidence indication. A simple out-of-distribution detection method, ODIN, is proposed in [52].…”