Self-Supervised Feature Augmentation for Large Image Object Detection

Lade...
Vorschaubild
Dateien
Zu diesem Dokument gibt es keine Dateien.
Datum
2020
Autor:innen
Dong, Weiming
Gu, Yang
Song, Zhichao
Meng, Yiping
Xu, Pengfei
Xu, Changsheng
Herausgeber:innen
Kontakt
ISSN der Zeitschrift
Electronic ISSN
ISBN
Bibliografische Daten
Verlag
Schriftenreihe
Auflagebezeichnung
URI (zitierfähiger Link)
ArXiv-ID
Internationale Patentnummer
Angaben zur Forschungsförderung
Projekt
Open Access-Veröffentlichung
Core Facility der Universität Konstanz
Gesperrt bis
Titel in einer weiteren Sprache
Forschungsvorhaben
Organisationseinheiten
Zeitschriftenheft
Publikationstyp
Zeitschriftenartikel
Publikationsstatus
Published
Erschienen in
IEEE Transactions on Image Processing. IEEE. 2020, 29, pp. 6745-6758. ISSN 1057-7149. eISSN 1941-0042. Available under: doi: 10.1109/TIP.2020.2993403
Zusammenfassung

Input scale plays an important role in modern detection frameworks, and an optimal training scale for images exists empirically. However, the optimal one usually cannot be reached in extremely large images under the memory constraint. In this study, we explore the scale effect inside the object detection pipeline and find that feature upsampling with the introduction of high-resolution information benefits the detection. Compared with direct input upscaling, feature upsampling trades a small performance loss for a large amount of memory savings. From these observations, we propose a self-supervised feature augmentation network, which takes downsampled images as inputs and aims to generate comparable features with the ones when feeding upscaled images to networks. We present a guided feature upsampling module, which takes downsampled images as inputs, to learn upscaled feature representations with the supervision of real large features acquired from upscaled images. In a self-supervised learning manner, we can introduce detailed information of images to the network. For an efficient feature upsampling, we design a residualized sub-pixel convolution block based on a sub-pixel convolution layer, which involves considerable information in upsampling process. Experiments on Mapillary Vistas Dataset (MVD), Cityscapes, and COCO are conducted to demonstrate the effectiveness of our method. On the MVD and Cityscapes detection benchmarks, in which the images are extremely large, our method surpasses current approaches. On COCO, the proposed method obtains comparable results to existing methods but with higher efficiency.

Zusammenfassung in einer weiteren Sprache
Fachgebiet (DDC)
004 Informatik
Schlagwörter
Konferenz
Rezension
undefined / . - undefined, undefined
Zitieren
ISO 690PAN, Xingjia, Fan TANG, Weiming DONG, Yang GU, Zhichao SONG, Yiping MENG, Pengfei XU, Oliver DEUSSEN, Changsheng XU, 2020. Self-Supervised Feature Augmentation for Large Image Object Detection. In: IEEE Transactions on Image Processing. IEEE. 2020, 29, pp. 6745-6758. ISSN 1057-7149. eISSN 1941-0042. Available under: doi: 10.1109/TIP.2020.2993403
BibTex
@article{Pan2020SelfS-49793,
  year={2020},
  doi={10.1109/TIP.2020.2993403},
  title={Self-Supervised Feature Augmentation for Large Image Object Detection},
  volume={29},
  issn={1057-7149},
  journal={IEEE Transactions on Image Processing},
  pages={6745--6758},
  author={Pan, Xingjia and Tang, Fan and Dong, Weiming and Gu, Yang and Song, Zhichao and Meng, Yiping and Xu, Pengfei and Deussen, Oliver and Xu, Changsheng}
}
RDF
<rdf:RDF
    xmlns:dcterms="http://purl.org/dc/terms/"
    xmlns:dc="http://purl.org/dc/elements/1.1/"
    xmlns:rdf="http://www.w3.org/1999/02/22-rdf-syntax-ns#"
    xmlns:bibo="http://purl.org/ontology/bibo/"
    xmlns:dspace="http://digital-repositories.org/ontologies/dspace/0.1.0#"
    xmlns:foaf="http://xmlns.com/foaf/0.1/"
    xmlns:void="http://rdfs.org/ns/void#"
    xmlns:xsd="http://www.w3.org/2001/XMLSchema#" > 
  <rdf:Description rdf:about="https://kops.uni-konstanz.de/server/rdf/resource/123456789/49793">
    <dc:contributor>Xu, Pengfei</dc:contributor>
    <dc:creator>Meng, Yiping</dc:creator>
    <dc:contributor>Song, Zhichao</dc:contributor>
    <foaf:homepage rdf:resource="http://localhost:8080/"/>
    <dc:creator>Gu, Yang</dc:creator>
    <dc:rights>terms-of-use</dc:rights>
    <dc:contributor>Pan, Xingjia</dc:contributor>
    <dcterms:abstract xml:lang="eng">Input scale plays an important role in modern detection frameworks, and an optimal training scale for images exists empirically. However, the optimal one usually cannot be reached in extremely large images under the memory constraint. In this study, we explore the scale effect inside the object detection pipeline and find that feature upsampling with the introduction of high-resolution information benefits the detection. Compared with direct input upscaling, feature upsampling trades a small performance loss for a large amount of memory savings. From these observations, we propose a self-supervised feature augmentation network, which takes downsampled images as inputs and aims to generate comparable features with the ones when feeding upscaled images to networks. We present a guided feature upsampling module, which takes downsampled images as inputs, to learn upscaled feature representations with the supervision of real large features acquired from upscaled images. In a self-supervised learning manner, we can introduce detailed information of images to the network. For an efficient feature upsampling, we design a residualized sub-pixel convolution block based on a sub-pixel convolution layer, which involves considerable information in upsampling process. Experiments on Mapillary Vistas Dataset (MVD), Cityscapes, and COCO are conducted to demonstrate the effectiveness of our method. On the MVD and Cityscapes detection benchmarks, in which the images are extremely large, our method surpasses current approaches. On COCO, the proposed method obtains comparable results to existing methods but with higher efficiency.</dcterms:abstract>
    <dc:contributor>Dong, Weiming</dc:contributor>
    <dc:creator>Pan, Xingjia</dc:creator>
    <dcterms:available rdf:datatype="http://www.w3.org/2001/XMLSchema#dateTime">2020-06-05T05:53:55Z</dcterms:available>
    <dcterms:rights rdf:resource="https://rightsstatements.org/page/InC/1.0/"/>
    <dc:contributor>Gu, Yang</dc:contributor>
    <dc:creator>Xu, Pengfei</dc:creator>
    <dc:contributor>Deussen, Oliver</dc:contributor>
    <dc:contributor>Tang, Fan</dc:contributor>
    <dc:creator>Deussen, Oliver</dc:creator>
    <dc:creator>Tang, Fan</dc:creator>
    <dc:creator>Song, Zhichao</dc:creator>
    <void:sparqlEndpoint rdf:resource="http://localhost/fuseki/dspace/sparql"/>
    <dc:date rdf:datatype="http://www.w3.org/2001/XMLSchema#dateTime">2020-06-05T05:53:55Z</dc:date>
    <dc:contributor>Meng, Yiping</dc:contributor>
    <dcterms:issued>2020</dcterms:issued>
    <dc:language>eng</dc:language>
    <dcterms:title>Self-Supervised Feature Augmentation for Large Image Object Detection</dcterms:title>
    <dc:creator>Xu, Changsheng</dc:creator>
    <dspace:isPartOfCollection rdf:resource="https://kops.uni-konstanz.de/server/rdf/resource/123456789/36"/>
    <dc:creator>Dong, Weiming</dc:creator>
    <dc:contributor>Xu, Changsheng</dc:contributor>
    <bibo:uri rdf:resource="https://kops.uni-konstanz.de/handle/123456789/49793"/>
    <dcterms:isPartOf rdf:resource="https://kops.uni-konstanz.de/server/rdf/resource/123456789/36"/>
  </rdf:Description>
</rdf:RDF>
Interner Vermerk
xmlui.Submission.submit.DescribeStep.inputForms.label.kops_note_fromSubmitter
Kontakt
URL der Originalveröffentl.
Prüfdatum der URL
Prüfungsdatum der Dissertation
Finanzierungsart
Kommentar zur Publikation
Allianzlizenz
Corresponding Authors der Uni Konstanz vorhanden
Internationale Co-Autor:innen
Universitätsbibliographie
Ja
Begutachtet
Ja
Diese Publikation teilen