Master Thesis Presentation

​Alfred Hazard och Filip Persson: Camera Pose Estimation and Multiview 2D to 3D Reconstruction

Abstract:
Generating 3D representations of objects is of interest in a wide breadth of industries. These representations are often created by hand through 3D modelling softwares such as CAD-derivatives. This in itself can be a complex process in order to capture desired detail and view dependent highlights. In this project, we investigate how Neural Radiance Fields (NeRF) can be used to extract the structure of an object within a bounded scene. NeRF encodes the structure of a scene in a Multi Layer Perceptron using a positional encoding heuristic. The training input is a 5-tuple consisting of a set of images and corresponding viewing directions of the cameras, and the output is the expected volume density and RGB color.
This project is a continuation of a Masters thesis by Isak Ernstig and David Olofsson, which has shown that accurate camera pose estimations are crucial to allow NeRF to render high fidelity views of a scene. In a feature poor environment traditional pose estimation pipelines using feature detection algorithms, such as the commonly used COLMAP estimator, have been shown to yield inadequate estimates. In this project, fiducial markers known as ArUco markers have been used to deduce accurate 2D-3D correspondences through detection and error correction. The detected markers allow for accurate usage of typical Computer Vision methodologies, such as Perspective-N-Point and Bundle Adjustment, allowing qualitative camera poses to be estimated. We conclude that our approach enables training a NeRF which may then subsequently render high fidelity novel views of real life objects. We also show the importance of accurate camera calibration and correct sampling intervals for rays when querying a trained NeRF model from a given viewpoint.

Supervisors: Erik Sintorn, Isak Ernstig, Ludwig Friborg och Oscar Andersson
Examiner: Mats Rudemo
Opponents: Kalle Bjurek och Victor Hagman
Category Student project presentation
Location: Euler, Skeppsgränd 3
Starts: 31 May, 2022, 15:15
Ends: 31 May, 2022, 16:45

Page manager Published: Wed 11 May 2022.