Scalable Full Flow with Learned Binary Descriptors

Gottfried Munda, Alexander Shekhovtsov, Patrick Knöbelreiter, Thomas Pock

Research output: Contribution to conferencePaperResearchpeer-review

Abstract

We propose a method for large displacement optical flow in
which local matching costs are learned by a convolutional neural network
(CNN) and a smoothness prior is imposed by a conditional random
field (CRF). We tackle the computation- and memory-intensive operations
on the 4D cost volume by a min-projection which reduces memory
complexity from quadratic to linear and binary descriptors for efficient
matching. This enables evaluation of the cost on the fly and allows to
perform learning and CRF inference on high resolution images without
ever storing the 4D cost volume. To address the problem of learning binary
descriptors we propose a new hybrid learning scheme. In contrast
to current state of the art approaches for learning binary CNNs we can
compute the exact non-zero gradient within our model. We compare several
methods for training binary descriptors and show results on public
available benchmarks.
Original languageEnglish
Publication statusPublished - 13 Sep 2017
Event39th German Conference on Pattern Recognition - Basel, Switzerland
Duration: 13 Sep 201615 Sep 2017
https://gcpr2017.dmi.unibas.ch/en/

Conference

Conference39th German Conference on Pattern Recognition
Abbreviated title GCPR 2017
CountrySwitzerland
CityBasel
Period13/09/1615/09/17
Internet address

Fingerprint

Costs
Optical flows
Image resolution
Neural networks
Data storage equipment

Cite this

Munda, G., Shekhovtsov, A., Knöbelreiter, P., & Pock, T. (2017). Scalable Full Flow with Learned Binary Descriptors. Paper presented at 39th German Conference on Pattern Recognition, Basel, Switzerland.

Scalable Full Flow with Learned Binary Descriptors. / Munda, Gottfried; Shekhovtsov, Alexander; Knöbelreiter, Patrick; Pock, Thomas.

2017. Paper presented at 39th German Conference on Pattern Recognition, Basel, Switzerland.

Research output: Contribution to conferencePaperResearchpeer-review

Munda, G, Shekhovtsov, A, Knöbelreiter, P & Pock, T 2017, 'Scalable Full Flow with Learned Binary Descriptors' Paper presented at 39th German Conference on Pattern Recognition, Basel, Switzerland, 13/09/16 - 15/09/17, .
Munda G, Shekhovtsov A, Knöbelreiter P, Pock T. Scalable Full Flow with Learned Binary Descriptors. 2017. Paper presented at 39th German Conference on Pattern Recognition, Basel, Switzerland.
Munda, Gottfried ; Shekhovtsov, Alexander ; Knöbelreiter, Patrick ; Pock, Thomas. / Scalable Full Flow with Learned Binary Descriptors. Paper presented at 39th German Conference on Pattern Recognition, Basel, Switzerland.
@conference{719f5dfb3aab41e786b48d6951cde53e,
title = "Scalable Full Flow with Learned Binary Descriptors",
abstract = "We propose a method for large displacement optical flow inwhich local matching costs are learned by a convolutional neural network(CNN) and a smoothness prior is imposed by a conditional randomfield (CRF). We tackle the computation- and memory-intensive operationson the 4D cost volume by a min-projection which reduces memorycomplexity from quadratic to linear and binary descriptors for efficientmatching. This enables evaluation of the cost on the fly and allows toperform learning and CRF inference on high resolution images withoutever storing the 4D cost volume. To address the problem of learning binarydescriptors we propose a new hybrid learning scheme. In contrastto current state of the art approaches for learning binary CNNs we cancompute the exact non-zero gradient within our model. We compare severalmethods for training binary descriptors and show results on publicavailable benchmarks.",
author = "Gottfried Munda and Alexander Shekhovtsov and Patrick Kn{\"o}belreiter and Thomas Pock",
year = "2017",
month = "9",
day = "13",
language = "English",
note = "39th German Conference on Pattern Recognition, GCPR 2017 ; Conference date: 13-09-2016 Through 15-09-2017",
url = "https://gcpr2017.dmi.unibas.ch/en/",

}

TY - CONF

T1 - Scalable Full Flow with Learned Binary Descriptors

AU - Munda, Gottfried

AU - Shekhovtsov, Alexander

AU - Knöbelreiter, Patrick

AU - Pock, Thomas

PY - 2017/9/13

Y1 - 2017/9/13

N2 - We propose a method for large displacement optical flow inwhich local matching costs are learned by a convolutional neural network(CNN) and a smoothness prior is imposed by a conditional randomfield (CRF). We tackle the computation- and memory-intensive operationson the 4D cost volume by a min-projection which reduces memorycomplexity from quadratic to linear and binary descriptors for efficientmatching. This enables evaluation of the cost on the fly and allows toperform learning and CRF inference on high resolution images withoutever storing the 4D cost volume. To address the problem of learning binarydescriptors we propose a new hybrid learning scheme. In contrastto current state of the art approaches for learning binary CNNs we cancompute the exact non-zero gradient within our model. We compare severalmethods for training binary descriptors and show results on publicavailable benchmarks.

AB - We propose a method for large displacement optical flow inwhich local matching costs are learned by a convolutional neural network(CNN) and a smoothness prior is imposed by a conditional randomfield (CRF). We tackle the computation- and memory-intensive operationson the 4D cost volume by a min-projection which reduces memorycomplexity from quadratic to linear and binary descriptors for efficientmatching. This enables evaluation of the cost on the fly and allows toperform learning and CRF inference on high resolution images withoutever storing the 4D cost volume. To address the problem of learning binarydescriptors we propose a new hybrid learning scheme. In contrastto current state of the art approaches for learning binary CNNs we cancompute the exact non-zero gradient within our model. We compare severalmethods for training binary descriptors and show results on publicavailable benchmarks.

UR - https://arxiv.org/pdf/1707.06427.pdf

M3 - Paper

ER -