2018/05/09 by Sergey Prokudin, Prokudin, Sergey, Peter Gehler +3 · 5 citations
Computer Science · Engineering · #Computer Vision and Pattern Recognition (cs.CV) #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences #Image and Object Detection Techniques #Robotics and Sensor-Based Localization
paper · pdf · doi:10.48550/arxiv.1805.03430
openalex publication_date 2018/05/09 · openalex created_date 2022/10/03 · openalex updated_date 2026/07/28
Modern deep learning systems successfully solve many perception tasks such as\nobject pose estimation when the input image is of high quality. However, in\nchallenging imaging conditions such as on low-resolution images or when the\nimage is corrupted by imaging artifacts, current systems degrade considerably\nin accuracy. While a loss in performance is unavoidable, we would like our\nmodels to quantify their uncertainty in order to achieve robustness against\nimages of varying quality. Probabilistic deep learning models combine the\nexpressive power of deep learning with uncertainty quantification. In this\npaper, we propose a novel probabilistic deep learning model for the task of\nangular regression. Our model uses von Mises distributions to predict a\ndistribution over object pose angle. Whereas a single von Mises distribution is\nmaking strong assumptions about the shape of the distribution, we extend the\nbasic model to predict a mixture of von Mises distributions. We show how to\nlearn a mixture model using a finite and infinite number of mixture components.\nOur model allows for likelihood-based training and efficient inference at test\ntime. We demonstrate on a number of challenging pose estimation datasets that\nour model produces calibrated probability predictions and competitive or\nsuperior point estimates compared to the current state-of-the-art.\n