Reconstruction or Retrieval? Investigating Neural 3D Reconstruction

Deep learning has become an increasingly popular approach for 3D reconstruction. While results look promising, deep learning produces black box networks which are hard to analyze. To apply such networks in the real world reliably, it is crucial that their functioning is well understood. In this pape...

Full description

Saved in:
Bibliographic Details
Main Authors: Gowda, Bhargava, Wunsche, Burkhard C., Shaw, Alex
Format: Conference Proceeding
Language:English
Subjects:
3D reconstruction
Deep learning
Image reconstruction
neural network
Reliability
Shape
shape retrieval
Three-dimensional displays
Training data
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Deep learning has become an increasingly popular approach for 3D reconstruction. While results look promising, deep learning produces black box networks which are hard to analyze. To apply such networks in the real world reliably, it is crucial that their functioning is well understood. In this paper we will evaluate to what extend current neural 3D reconstruction algorithms perform true 3D reconstruction, i.e. whether they are able to produce novel 3D content not contained in the training data set, or whether the networks are just performing some modified shape retrieval. The transference of reconstruction ability from multi view to single view reconstruction is also tested. Our results suggest that neural 3D reconstruction algorithms can indeed learn to perform reconstruction but the transference of this ability between single and multi view is not well supported.
ISSN:2151-2205
DOI:10.1109/IVCNZ54163.2021.9653224