Novelty detection in very high resolution urban scenes with Density Forests

Cyril Wendl, Diego Marcos, Devis Tuia

Research output: Chapter in Book/Report/Conference proceedingConference paper

Abstract

Uncertainty in deep learning has recently received a lot of attention. While deep neural networks have shown better accuracy than other competing methods in many benchmarks, it has been shown that they may yield wrong predictions with unreasonably high confidence. This has increased the interest in methods that help providing better confidence estimates in neural networks, some using specifically designed architectures with probabilistic building blocks, and others using a standard architecture with an additional confidence estimation step based on its output. This work proposes a confidence estimation method for Convolutional Neural Networks based on fitting a forest of randomized density estimation decision trees to the network activations before the final classification layer and compares it to other confidence estimation methods based on standard architectures. The methods are compared on a semantic labelling dataset with very high resolution satellite imagery. Our results show that methods based on intermediate network activations lead to better confidence estimates in novelty detection, i.e., in the discovery of classes that are not present in the training set.

Original languageEnglish
Title of host publication2019 Joint Urban Remote Sensing Event, JURSE 2019
PublisherInstitute of Electrical and Electronics Engineers Inc.
ISBN (Electronic)9781728100098, 9781728100081
ISBN (Print)9781728100104
DOIs
Publication statusPublished - 22 Aug 2019
Event2019 Joint Urban Remote Sensing Event, JURSE 2019 - Vannes, France
Duration: 22 May 201924 May 2019

Publication series

NameJoint Urban Remote Sensing Event (JURSE)
PublisherIEEE
ISSN (Print)2334-0932
ISSN (Electronic)2642-9535

Conference

Conference2019 Joint Urban Remote Sensing Event, JURSE 2019
CountryFrance
CityVannes
Period22/05/1924/05/19

Fingerprint

confidence
high resolution
estimation method
neural network
Chemical activation
Neural networks
activation
Satellite imagery
Decision trees
satellite imagery
Labeling
semantics
learning
Semantics
estimates
marking
detection
method
education
prediction

Keywords

  • Convolutional Neural Networks
  • Density Forest
  • land cover
  • novelty detection
  • Uncertainty

Cite this

Wendl, C., Marcos, D., & Tuia, D. (2019). Novelty detection in very high resolution urban scenes with Density Forests. In 2019 Joint Urban Remote Sensing Event, JURSE 2019 [8808974] (Joint Urban Remote Sensing Event (JURSE)). Institute of Electrical and Electronics Engineers Inc.. https://doi.org/10.1109/JURSE.2019.8808974
Wendl, Cyril ; Marcos, Diego ; Tuia, Devis. / Novelty detection in very high resolution urban scenes with Density Forests. 2019 Joint Urban Remote Sensing Event, JURSE 2019. Institute of Electrical and Electronics Engineers Inc., 2019. (Joint Urban Remote Sensing Event (JURSE)).
@inproceedings{7adf15e9ee774257aa02b8f55a9d8aec,
title = "Novelty detection in very high resolution urban scenes with Density Forests",
abstract = "Uncertainty in deep learning has recently received a lot of attention. While deep neural networks have shown better accuracy than other competing methods in many benchmarks, it has been shown that they may yield wrong predictions with unreasonably high confidence. This has increased the interest in methods that help providing better confidence estimates in neural networks, some using specifically designed architectures with probabilistic building blocks, and others using a standard architecture with an additional confidence estimation step based on its output. This work proposes a confidence estimation method for Convolutional Neural Networks based on fitting a forest of randomized density estimation decision trees to the network activations before the final classification layer and compares it to other confidence estimation methods based on standard architectures. The methods are compared on a semantic labelling dataset with very high resolution satellite imagery. Our results show that methods based on intermediate network activations lead to better confidence estimates in novelty detection, i.e., in the discovery of classes that are not present in the training set.",
keywords = "Convolutional Neural Networks, Density Forest, land cover, novelty detection, Uncertainty",
author = "Cyril Wendl and Diego Marcos and Devis Tuia",
year = "2019",
month = "8",
day = "22",
doi = "10.1109/JURSE.2019.8808974",
language = "English",
isbn = "9781728100104",
series = "Joint Urban Remote Sensing Event (JURSE)",
publisher = "Institute of Electrical and Electronics Engineers Inc.",
booktitle = "2019 Joint Urban Remote Sensing Event, JURSE 2019",

}

Wendl, C, Marcos, D & Tuia, D 2019, Novelty detection in very high resolution urban scenes with Density Forests. in 2019 Joint Urban Remote Sensing Event, JURSE 2019., 8808974, Joint Urban Remote Sensing Event (JURSE), Institute of Electrical and Electronics Engineers Inc., 2019 Joint Urban Remote Sensing Event, JURSE 2019, Vannes, France, 22/05/19. https://doi.org/10.1109/JURSE.2019.8808974

Novelty detection in very high resolution urban scenes with Density Forests. / Wendl, Cyril; Marcos, Diego; Tuia, Devis.

2019 Joint Urban Remote Sensing Event, JURSE 2019. Institute of Electrical and Electronics Engineers Inc., 2019. 8808974 (Joint Urban Remote Sensing Event (JURSE)).

Research output: Chapter in Book/Report/Conference proceedingConference paper

TY - GEN

T1 - Novelty detection in very high resolution urban scenes with Density Forests

AU - Wendl, Cyril

AU - Marcos, Diego

AU - Tuia, Devis

PY - 2019/8/22

Y1 - 2019/8/22

N2 - Uncertainty in deep learning has recently received a lot of attention. While deep neural networks have shown better accuracy than other competing methods in many benchmarks, it has been shown that they may yield wrong predictions with unreasonably high confidence. This has increased the interest in methods that help providing better confidence estimates in neural networks, some using specifically designed architectures with probabilistic building blocks, and others using a standard architecture with an additional confidence estimation step based on its output. This work proposes a confidence estimation method for Convolutional Neural Networks based on fitting a forest of randomized density estimation decision trees to the network activations before the final classification layer and compares it to other confidence estimation methods based on standard architectures. The methods are compared on a semantic labelling dataset with very high resolution satellite imagery. Our results show that methods based on intermediate network activations lead to better confidence estimates in novelty detection, i.e., in the discovery of classes that are not present in the training set.

AB - Uncertainty in deep learning has recently received a lot of attention. While deep neural networks have shown better accuracy than other competing methods in many benchmarks, it has been shown that they may yield wrong predictions with unreasonably high confidence. This has increased the interest in methods that help providing better confidence estimates in neural networks, some using specifically designed architectures with probabilistic building blocks, and others using a standard architecture with an additional confidence estimation step based on its output. This work proposes a confidence estimation method for Convolutional Neural Networks based on fitting a forest of randomized density estimation decision trees to the network activations before the final classification layer and compares it to other confidence estimation methods based on standard architectures. The methods are compared on a semantic labelling dataset with very high resolution satellite imagery. Our results show that methods based on intermediate network activations lead to better confidence estimates in novelty detection, i.e., in the discovery of classes that are not present in the training set.

KW - Convolutional Neural Networks

KW - Density Forest

KW - land cover

KW - novelty detection

KW - Uncertainty

U2 - 10.1109/JURSE.2019.8808974

DO - 10.1109/JURSE.2019.8808974

M3 - Conference paper

SN - 9781728100104

T3 - Joint Urban Remote Sensing Event (JURSE)

BT - 2019 Joint Urban Remote Sensing Event, JURSE 2019

PB - Institute of Electrical and Electronics Engineers Inc.

ER -

Wendl C, Marcos D, Tuia D. Novelty detection in very high resolution urban scenes with Density Forests. In 2019 Joint Urban Remote Sensing Event, JURSE 2019. Institute of Electrical and Electronics Engineers Inc. 2019. 8808974. (Joint Urban Remote Sensing Event (JURSE)). https://doi.org/10.1109/JURSE.2019.8808974