Please use this identifier to cite or link to this item:
|Title:||Improvements to Multi-View Stereo|
|Abstract:||3D reconstruction is the problem of using a set of images of an object, place, person, thing, etc.to create a 3D reconstruction of the subject of the images. A method of 3D reconstruction is Multi-view stereo (MVS) which refers to a set of techniques which use stereo correspondence among a large number of images as a cue for 3D reconstruction. I further improve the state of the art approaches which use this route in order to perform better on the DTU Robot Image MVS benchmark for the 3D reconstruction task by improving on the model UCSNet [green12] by providing three novel contributions: 1) introduce surface normals into the loss function to help the network better learn semantic information, 2) modify the cost volume construction of the network, and 3) consider geometric consistency when training the network. Specifically, I have found that adding surface normals to the loss function improves the completion of the model leads to a 4% more complete reconstruction while only reducing accuracy by less than 1%. Modifying the cost volume construction however did not majorly change the model. Furthermore, adding geometric consistency information improves the accuracy of the network 6.9% and however reduces completion of 7%. In conclusion,we are able to improve upon the current state of the art UCSNet [green12] through using surface normal supervision.|
|Type of Material:||Princeton University Senior Theses|
|Appears in Collections:||Computer Science, 1988-2021|
Files in This Item:
|PALISKARA-SAHAN-THESIS.pdf||4.79 MB||Adobe PDF||Request a copy|
Items in Dataspace are protected by copyright, with all rights reserved, unless otherwise indicated.