COMPARISON OF UNCERTAINTY QUANTIFICATION METHODS FOR CNN-BASED REGRESSION
- Karlsruhe Institute of Technology (KIT), Institute of Photogrammetry and Remote Sensing (IPF), Karlsruhe, Germany
Keywords: Deep Learning, Regression, Convolutional Neural Network, Uncertainty Quantification, Bayesian Modelling, Variational Inference
Abstract. The evaluation of reliability is not only of high importance for safety-critical deep learning applications but for object pose estimation as well. The uncertainty of the result is one way to express its reliability. In order to better understand existing uncertainty quantification (UQ) methods and their performance on image-based regression tasks, we use a small CNN and various scenarios to evaluate the estimated uncertainties. The evaluation is done on different simplistic synthetic datasets, consisting of gray-scale images of squares on a darker background. We train the CNN to predict the square center position of the square in the image. We compare how different UQ methods perform under dataset shift, rotation, occlusion, noise changes in the images.