Prediction Surface Uncertainty Quantification in Object Detection Models for Autonomous Driving

Object detection in autonomous cars is commonly based on camera images and\nLidar inputs, which are often used to train prediction models such as deep\nartificial neural networks for decision making for object recognition,\nadjusting speed, etc. A mistake in such decision making can be damaging; thus,\nit is vital to measure the reliability of decisions made by such prediction\nmodels via uncertainty measurement. Uncertainty, in deep learning models, is\noften measured for classification problems. However, deep learning models in\nautonomous driving are often multi-output regression models. Hence, we propose\na novel method called PURE (Prediction sURface uncErtainty) for measuring\nprediction uncertainty of such regression models. We formulate the object\nrecognition problem as a regression model with more than one outputs for\nfinding object locations in a 2-dimensional camera view. For evaluation, we\nmodified three widely-applied object recognition models (i.e., YoLo, SSD300 and\nSSD512) and used the KITTI, Stanford Cars, Berkeley DeepDrive, and NEXET\ndatasets. Results showed the statistically significant negative correlation\nbetween prediction surface uncertainty and prediction accuracy suggesting that\nuncertainty significantly impacts the decisions made by autonomous driving.\n