Seediscussions,stats,andauthorprofilesforthispublicationat:https://www.researchgate.net/publication/278027951

place-recognition-wxbs-poster

Data·June2015

CITATIONS

0

READS

18

3authors:

Someoftheauthorsofthispublicationarealsoworkingontheserelatedprojects:

VisualtrackingevaluationViewproject

machinelearningViewproject

DmytroMishkin

CzechTechnicalUniversityinPrague

24PUBLICATIONS93CITATIONS

SEEPROFILE

MichalPerdoch

CzechTechnicalUniversityinPrague

26PUBLICATIONS722CITATIONS

SEEPROFILE

JiriMatas

CzechTechnicalUniversityinPrague

271PUBLICATIONS21,780CITATIONS

SEEPROFILE

AllcontentfollowingthispagewasuploadedbyDmytroMishkinon11June2015.

Theuserhasrequestedenhancementofthedownloadedfile.

IEEE 2015 Conference on

Computer Vision and Pattern

Recognition Place Recognition with WxBS Retrieval

Dmytro Mishkin, Michal Perdoch, Jiri Matas Center for Machine Perception, Czech Technical University in Prague

Yes, no bridge in winter yet

Workshop on Visual Place

Recognition in Changing

Environments

We present a novel visual place recognition method designed to operate in challenging conditions such as encountered in day to night or winter to summer matching. The proposed WxBS Retrieval method is novel in enriching a bag of words approach with the use of multiple detectors, descriptors with suitable visual vocabularies, view synthesis, and adaptive thresholding to compensate for large variations in contrast and richness of features in different conditions.

Adjacency Motion Model Structure of WxBS Retrieval and Rescoring

The authors were supported by the Czech Science Foundation Project GACR P103/12/G084, the Technology Agency of the Czech Republic research program TE01020415 (V3C -- Visual Computing Competence Center) and by the MSMT LL1303 ERC-CZ grant.

WxBS-MODS Matching and RANSAC Verification

Abstract

Contributions

WxBS-MODS Feature Detection & Description

References

• Retrieval database augmentation by affine view synthesis

• Use both RootSIFT and HalfRootSIFT for image retrieval

• WxBS-MODS matcher for validation

Advantages

• Efficiency: same features for retrieval and matching verification

• Easy implementable in standard frameworks

• High recall & precision even with big appearance and geometric changes

VPRiCE Challenge 2015 Results

query

output

No match

query

output

• Retrieval results rescoring by iterative two view matching with MODS [2], using the same features as in retrieval.

[1] D. Mishkin, J. Matas, M. Perdoch, and K. Lenc. WxBS: Wide Baseline Stereo Generalizations. CoRR, abs/1504.06603, Apr. 2015 [2] D. Mishkin, M. Perdoch, and J. Matas. Mods: Fast and robust method for two-view matching. CoRR, abs/1503.02619, Mar. 2015 [3] J. Chen, J. Tian, N. Lee, J. Zheng, R. Smith, and A. Laine. A partial intensity invariant feature descriptor for multimodal retinal image registration. Biomedical Engineering, IEEE Transactions on, 57(7):1707–1718, 2010

Method Prec Recall F1

MAPIR (CNN) 0.747 0.836 0.789

Bonn (CNN) 0.726 0.758 0.741

BoW HalfRootSIFT 0.530 0.890 0.665

BoW Half & RootSIFT 0.538 1.000* 0.700

BoW Half & RootSIFT & MODS + adj. model

0.821 0.825 0.823

• Standard TF-IDF inverted file retrieval engine, 4 vocabularies: for RootSIFT and HalfRootSIFT and for MSERs and Hessian

• Spatial verification, with a short-list length = 1K images • Label soft-assignment:

2 labels per words, alternative labels treated independently • No query expansion for efficiency reasons

BoW engine

Online stage

Offline stage

HalfRootSIFT

HalfSIFT bins SIFT bins

ground truth No match No match

Exploits the information about linear ordering of query reference images: 1. Initialization: For 3 consecutive query images select adjacency

model that is consistent with retrieved reference images. If none is found, output “no match” and repeat init. for the next image.

2. Propagation: Check if the next query image is MODS

matching with the next reference image according to adjacency model, if not re-initialize with step 1.

query reference

Fast move

Move

No move

No match

* Results are shown as reported by VPRICE challenge organizers, with ±1 frame positional error tolerance. We believe than recall = 1.0 is an artifact of results calculation when the method does not output “no match“ answer.

HalfSIFT [3] is a modification of SIFT Robust to local contrast reversal, that maps gradient orientaton from 0-360°to 0-180°

View publication statsView publication stats