Object Detection in Equirectangular Panorama

Wenyan Yang; Yanlin Qian; Joni-Kristian Kämäräinen; Francesco Cricri; Lixin Fan

doi:10.1109/ICPR.2018.8546070

Object Detection in Equirectangular Panorama

Wenyan Yang, Yanlin Qian, Joni-Kristian Kämäräinen, Francesco Cricri, Lixin Fan

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › Scientific › peer-review

87 Citations (Scopus)

Abstract

We introduce a high-resolution equirectangular panorama (aka 360-degree, virtual reality, VR) dataset for object detection and propose a multi-projection variant of the YOLO detector. The main challenges with equirectangular panorama images are i) the lack of annotated training data, ii) high-resolution imagery and iii) severe geometric distortions of objects near the panorama projection poles. In this work, we solve the challenges by I) using training examples available in the “conventional datasets” (ImageNet and COCO), II) employing only low resolution images that require only moderate GPU computing power and memory, and III) our multi-projection YOLO handles projection distortions by making multiple stereographic sub-projections. In our experiments, YOLO outperforms the other state-of-the-art detector, Faster R-CNN, and our multi-projection YOLO achieves the best accuracy with low-resolution input.

Original language	English
Title of host publication	2018 24th International Conference on Pattern Recognition (ICPR)
Publisher	IEEE
Pages	2190-2195
Number of pages	6
ISBN (Electronic)	978-1-5386-3788-3
ISBN (Print)	978-1-5386-3789-0
DOIs	https://doi.org/10.1109/ICPR.2018.8546070
Publication status	Published - Aug 2018
Publication type	A4 Article in conference proceedings
Event	International Conference on Pattern Recognition - Duration: 1 Jan 1900 → …

Publication series

Name
ISSN (Print)	1051-4651

Conference

Conference	International Conference on Pattern Recognition
Period	1/01/00 → …

Keywords

image resolution
neural nets
object detection
virtual reality
YOLO detector
equirectangular panorama images
annotated training data
high-resolution imagery
panorama projection poles
training examples
conventional datasets
low resolution images
multiprojection YOLO
projection distortions
multiple stereographic sub-projections
low-resolution input
high-resolution equirectangular panorama dataset
geometric distortions
moderate GPU computing power
CNN
Detectors
Distortion
Object detection
Cameras
Image resolution
Virtual reality
Graphics processing units

Publication forum classification

Publication forum level 1

Access to Document

10.1109/ICPR.2018.8546070

Cite this

@inproceedings{84ffa60d2015464b817900ada7443ca4,

title = "Object Detection in Equirectangular Panorama",

abstract = "We introduce a high-resolution equirectangular panorama (aka 360-degree, virtual reality, VR) dataset for object detection and propose a multi-projection variant of the YOLO detector. The main challenges with equirectangular panorama images are i) the lack of annotated training data, ii) high-resolution imagery and iii) severe geometric distortions of objects near the panorama projection poles. In this work, we solve the challenges by I) using training examples available in the “conventional datasets” (ImageNet and COCO), II) employing only low resolution images that require only moderate GPU computing power and memory, and III) our multi-projection YOLO handles projection distortions by making multiple stereographic sub-projections. In our experiments, YOLO outperforms the other state-of-the-art detector, Faster R-CNN, and our multi-projection YOLO achieves the best accuracy with low-resolution input.",

keywords = "image resolution, neural nets, object detection, virtual reality, YOLO detector, equirectangular panorama images, annotated training data, high-resolution imagery, panorama projection poles, training examples, conventional datasets, low resolution images, multiprojection YOLO, projection distortions, multiple stereographic sub-projections, low-resolution input, high-resolution equirectangular panorama dataset, geometric distortions, moderate GPU computing power, CNN, Detectors, Distortion, Object detection, Cameras, Image resolution, Virtual reality, Graphics processing units",

author = "Wenyan Yang and Yanlin Qian and Joni-Kristian K{\"a}m{\"a}r{\"a}inen and Francesco Cricri and Lixin Fan",

note = "EXT={"}Cricri, Francesco{"}; International Conference on Pattern Recognition ; Conference date: 01-01-1900",

year = "2018",

month = aug,

doi = "10.1109/ICPR.2018.8546070",

language = "English",

isbn = "978-1-5386-3789-0",

publisher = "IEEE",

pages = "2190--2195",

booktitle = "2018 24th International Conference on Pattern Recognition (ICPR)",

address = "United States",

}

TY - GEN

T1 - Object Detection in Equirectangular Panorama

AU - Yang, Wenyan

AU - Qian, Yanlin

AU - Kämäräinen, Joni-Kristian

AU - Cricri, Francesco

AU - Fan, Lixin

N1 - EXT="Cricri, Francesco"

PY - 2018/8

Y1 - 2018/8

N2 - We introduce a high-resolution equirectangular panorama (aka 360-degree, virtual reality, VR) dataset for object detection and propose a multi-projection variant of the YOLO detector. The main challenges with equirectangular panorama images are i) the lack of annotated training data, ii) high-resolution imagery and iii) severe geometric distortions of objects near the panorama projection poles. In this work, we solve the challenges by I) using training examples available in the “conventional datasets” (ImageNet and COCO), II) employing only low resolution images that require only moderate GPU computing power and memory, and III) our multi-projection YOLO handles projection distortions by making multiple stereographic sub-projections. In our experiments, YOLO outperforms the other state-of-the-art detector, Faster R-CNN, and our multi-projection YOLO achieves the best accuracy with low-resolution input.

AB - We introduce a high-resolution equirectangular panorama (aka 360-degree, virtual reality, VR) dataset for object detection and propose a multi-projection variant of the YOLO detector. The main challenges with equirectangular panorama images are i) the lack of annotated training data, ii) high-resolution imagery and iii) severe geometric distortions of objects near the panorama projection poles. In this work, we solve the challenges by I) using training examples available in the “conventional datasets” (ImageNet and COCO), II) employing only low resolution images that require only moderate GPU computing power and memory, and III) our multi-projection YOLO handles projection distortions by making multiple stereographic sub-projections. In our experiments, YOLO outperforms the other state-of-the-art detector, Faster R-CNN, and our multi-projection YOLO achieves the best accuracy with low-resolution input.

KW - image resolution

KW - neural nets

KW - object detection

KW - virtual reality

KW - YOLO detector

KW - equirectangular panorama images

KW - annotated training data

KW - high-resolution imagery

KW - panorama projection poles

KW - training examples

KW - conventional datasets

KW - low resolution images

KW - multiprojection YOLO

KW - projection distortions

KW - multiple stereographic sub-projections

KW - low-resolution input

KW - high-resolution equirectangular panorama dataset

KW - geometric distortions

KW - moderate GPU computing power

KW - CNN

KW - Detectors

KW - Distortion

KW - Object detection

KW - Cameras

KW - Image resolution

KW - Virtual reality

KW - Graphics processing units

U2 - 10.1109/ICPR.2018.8546070

DO - 10.1109/ICPR.2018.8546070

M3 - Conference contribution

SN - 978-1-5386-3789-0

SP - 2190

EP - 2195

BT - 2018 24th International Conference on Pattern Recognition (ICPR)

PB - IEEE

T2 - International Conference on Pattern Recognition

Y2 - 1 January 1900

ER -

Object Detection in Equirectangular Panorama

Abstract

Publication series

Conference

Keywords

Publication forum classification

Access to Document

Fingerprint

Cite this