KIT | KIT-Bibliothek | Impressum | Datenschutz

Activity-aware attributes for zero-shot driver behavior recognition

Reiß, Simon; Roitberg, Alina; Haurilet, Monica; Stiefelhagen, Rainer

Abstract (englisch):
In real-world environments, such as the vehicle cabin, we have to deal with novel concepts as they arise. To this end, we introduce ZS-Drive&Act - the first zero-shot activity classification benchmark specifically aimed at recognizing previously unseen driver behaviors. ZS-Drive&Act is unique due to its focus on fine-grained activities and presence of activity-driven attributes, which are automatically derived from a hierarchical annotation scheme. We adopt and evaluate multiple off-the-shelf zero-shot learning methods on our benchmark, showcasing the difficulties of such models when moving to our application-specific task. We further extend the prominent method based on feature generating Wasserstein GANs with a fusion strategy for linking semantic attributes and word vectors representing the behavior labels. Our experiments demonstrate the effectiveness of leveraging both semantic spaces simultaneously, improving the recognition rate by 2.79%.



Originalveröffentlichung
DOI: 10.1109/CVPRW50498.2020.00459
Scopus
Zitationen: 2
Dimensions
Zitationen: 1
Zugehörige Institution(en) am KIT Institut für Anthropomatik und Robotik (IAR)
Studienzentrum für Sehgeschädigte (SZS)
Publikationstyp Proceedingsbeitrag
Publikationsjahr 2020
Sprache Englisch
Identifikator ISBN: 978-1-72819-360-1
ISSN: 2160-7508
KITopen-ID: 1000124084
Erschienen in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, CVPRW 2020; Virtual, Online; United States; 14 June 2020 through 19 June 2020
Verlag Institute of Electrical and Electronics Engineers (IEEE)
Seiten 3950-3955
Serie IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops ; 2020-June
Vorab online veröffentlicht am 28.07.2020
Nachgewiesen in Dimensions
Scopus
KIT – Die Forschungsuniversität in der Helmholtz-Gemeinschaft
KITopen Landing Page