Self-Supervised Learning for Stereo Reconstruction on Aerial Images

Research output: Contribution to conferencePaper

Abstract

Recent developments established deep learning as an inevitable tool to boost the performance of dense matching and stereo estimation. On the downside, learning these networks requires a substantial amount of training data to be successful. Consequently, the application of these models outside of the laboratory is far from straight forward. In this work we propose a self-supervised training procedure that allows us
to adapt our network to the specific (imaging) characteristics of the dataset at hand, without the requirement of external ground truth data. We instead generate interim training data by running our intermediate network on the whole dataset, followed by conservative outlier filtering. Bootstrapped from a pre-trained version of our hybrid CNN-CRF model, we alternate the generation of training data and network training.
With this simple concept we are able to lift the completeness and accuracy of the pretrained version significantly. We also show that our final model compares favorably to other popular stereo estimation algorithms on an aerial dataset.

Conference

Conference38th International Geoscience and Remote Sensing Symposium
Abbreviated titleIGARSS
CountrySpain
CityValencia
Period22/07/1827/07/18
Internet address

Fingerprint

Supervised learning
Antennas
Imaging techniques

Keywords

  • large scale 3D
  • dense matching
  • CNN

Cite this

Knöbelreiter, P., Vogel, C., & Pock, T. (2018). Self-Supervised Learning for Stereo Reconstruction on Aerial Images. Paper presented at 38th International Geoscience and Remote Sensing Symposium, Valencia, Spain.

Self-Supervised Learning for Stereo Reconstruction on Aerial Images. / Knöbelreiter, Patrick; Vogel, Christoph; Pock, Thomas.

2018. Paper presented at 38th International Geoscience and Remote Sensing Symposium, Valencia, Spain.

Research output: Contribution to conferencePaper

Knöbelreiter, P, Vogel, C & Pock, T 2018, 'Self-Supervised Learning for Stereo Reconstruction on Aerial Images' Paper presented at 38th International Geoscience and Remote Sensing Symposium, Valencia, Spain, 22/07/18 - 27/07/18, .
Knöbelreiter P, Vogel C, Pock T. Self-Supervised Learning for Stereo Reconstruction on Aerial Images. 2018. Paper presented at 38th International Geoscience and Remote Sensing Symposium, Valencia, Spain.
Knöbelreiter, Patrick ; Vogel, Christoph ; Pock, Thomas. / Self-Supervised Learning for Stereo Reconstruction on Aerial Images. Paper presented at 38th International Geoscience and Remote Sensing Symposium, Valencia, Spain.
@conference{efb980459aeb483e974dcf51a8920c93,
title = "Self-Supervised Learning for Stereo Reconstruction on Aerial Images",
abstract = "Recent developments established deep learning as an inevitable tool to boost the performance of dense matching and stereo estimation. On the downside, learning these networks requires a substantial amount of training data to be successful. Consequently, the application of these models outside of the laboratory is far from straight forward. In this work we propose a self-supervised training procedure that allows usto adapt our network to the specific (imaging) characteristics of the dataset at hand, without the requirement of external ground truth data. We instead generate interim training data by running our intermediate network on the whole dataset, followed by conservative outlier filtering. Bootstrapped from a pre-trained version of our hybrid CNN-CRF model, we alternate the generation of training data and network training.With this simple concept we are able to lift the completeness and accuracy of the pretrained version significantly. We also show that our final model compares favorably to other popular stereo estimation algorithms on an aerial dataset.",
keywords = "large scale 3D, dense matching, CNN",
author = "Patrick Kn{\"o}belreiter and Christoph Vogel and Thomas Pock",
year = "2018",
month = "7",
day = "22",
language = "English",
note = "38th International Geoscience and Remote Sensing Symposium, IGARSS ; Conference date: 22-07-2018 Through 27-07-2018",
url = "https://igarss2018.org",

}

TY - CONF

T1 - Self-Supervised Learning for Stereo Reconstruction on Aerial Images

AU - Knöbelreiter,Patrick

AU - Vogel,Christoph

AU - Pock,Thomas

PY - 2018/7/22

Y1 - 2018/7/22

N2 - Recent developments established deep learning as an inevitable tool to boost the performance of dense matching and stereo estimation. On the downside, learning these networks requires a substantial amount of training data to be successful. Consequently, the application of these models outside of the laboratory is far from straight forward. In this work we propose a self-supervised training procedure that allows usto adapt our network to the specific (imaging) characteristics of the dataset at hand, without the requirement of external ground truth data. We instead generate interim training data by running our intermediate network on the whole dataset, followed by conservative outlier filtering. Bootstrapped from a pre-trained version of our hybrid CNN-CRF model, we alternate the generation of training data and network training.With this simple concept we are able to lift the completeness and accuracy of the pretrained version significantly. We also show that our final model compares favorably to other popular stereo estimation algorithms on an aerial dataset.

AB - Recent developments established deep learning as an inevitable tool to boost the performance of dense matching and stereo estimation. On the downside, learning these networks requires a substantial amount of training data to be successful. Consequently, the application of these models outside of the laboratory is far from straight forward. In this work we propose a self-supervised training procedure that allows usto adapt our network to the specific (imaging) characteristics of the dataset at hand, without the requirement of external ground truth data. We instead generate interim training data by running our intermediate network on the whole dataset, followed by conservative outlier filtering. Bootstrapped from a pre-trained version of our hybrid CNN-CRF model, we alternate the generation of training data and network training.With this simple concept we are able to lift the completeness and accuracy of the pretrained version significantly. We also show that our final model compares favorably to other popular stereo estimation algorithms on an aerial dataset.

KW - large scale 3D

KW - dense matching

KW - CNN

M3 - Paper

ER -