Unveiling the Depths: Enhancing Surface Reconstruction from Satellite Images, using Neural Kernel Fields
2024 (English)Independent thesis Advanced level (degree of Master (Two Years)), 28 HE credits
Student thesis
Abstract [en]
Satellite imaging is an essential tool for modern remote sensing that provides valuable insights into our dynamic world. Maxar is at the forefront of technological innovation in capturing and analyzing satellite imagery. However, the current process of reconstructing terrain features from satellite data has some limitations, particularly in accurately representing complex three-dimensional structures such as bridges and subterranean features.
This thesis aims to utilize deep learning algorithms to tackle the challenges of capturing intricate terrain features and improving the accuracy of the reconstructed surfaces. More specifically, deep learning surface reconstruction was evaluated through Neural Kernel Surface Reconstruction (NKSR).
The use of NKSR does seem to improve the accuracy of the mesh in the same city as the one it was trained on and did improve the mesh quality on flat surfaces. It also seems that adding more features positively affects the quality of the mesh, but it is clear that a mesh created using trained NKSR usually has more holes and free-flowing noise than the original Maxar mesh.
Place, publisher, year, edition, pages
2024. , p. 61
Keywords [en]
NKSR, Surface Reconstruction, Point Cloud, Satellite Imagery, Machine Learning, Deep Learning
National Category
Computer graphics and computer vision
Identifiers
URN: urn:nbn:se:liu:diva-205860OAI: oai:DiVA.org:liu-205860DiVA, id: diva2:1882588
External cooperation
Maxar Intelligence
Subject / course
Computer Vision Laboratory
Supervisors
Examiners
2024-08-122024-07-052025-02-07Bibliographically approved