Evaluating uncertainty estimation techniques using deep leaning models for classification of point clouds

Introduction and goal of the work

Deep learning models have proven their ability to perform powerful classification and segmentation considering both 2d and 3D datasets. However, these classifications are often assumed to be accurate, without providing how reliable are these decisions made by the model. There are some possible techniques to convert a stochastic deep learning model into a probabilistic one, to provide a distribution over the weights, instead of finding point estimation for them, leading to an output distribution. Such distribution over output of a model shows not only the true class but also its reliability as the variance of the distribution. 

Tasks

1. Considering a suitable deep learning model for 3D point cloud classification, like KPConv, PointNet, ...
2. Considering an uncertainty estimation algorithm to estimate the uncertainty of classification from task 1, including Dropout-based uncertainty estimation, Bayesian neural networks, Direct Modelling methods, Deep Ensemble methods
3.  Conducting a literature review on evaluation techniques of uncertainty estimation 
4. Implementing evaluation techniques to evaluate task 2

Resources 

►    Point cloud data from LiDAR
►    Benchmark dataset like SemanticKitti

Requirements

►    Proficient programming skills in Python
►    Experience in Deep Leaning 
►    Experience in working with point clouds 
►    Understanding of probabilistic Neural Networks is an advantage

Contact Person 

M.Sc. Hanieh Shojaei (E-Mail hanieh.shojaei@ikg.uni-hannover.de, Tel. 762-2472)
Institut für Kartographie und Geoinformatik, Appelstraße 9 a, 30167 Hannover, Raum 616