Paper accepted in ICMR 2020

Asra Aslam and Edward Curry, “Object Detection for Unseen Domains while Reducing Response Time using Knowledge Transfer in Multimedia Event Processing.” Proceedings of the 2020 ACM on International Conference on Multimedia Retrieval (ICMR).

Abstract

Event recognition is among one of the popular areas of smart citiesthat has attracted great attention for researchers. Since Internet ofThings (IoT) is mainly focused on scalar data events, research isshifting towards the Internet of Multimedia Things (IoMT) and isstill in infancy. Presently multimedia event-based solutions providelow response-time, but they are domain-specic and can handleonly familiar classes (bounded vocabulary). However multiple ap-plications within smart cities may require processing of numerousfamiliar as well as unseen concepts (unbounded vocabulary) in theform of subscriptions. Deep neural network-based techniques arepopular for image recognition, but have the limitation of trainingof classiers for unseen concepts as well as the requirement of an-notated bounding boxes with images. In this work, we explore theproblem of training of classiers for unseen/unknown classes whilereducing response-time of multimedia event processing (specicallyobject detection). We proposed two domain adaptation based mod-els while leveraging Transfer Learning (TL) and Large Scale Detec-tion through Adaptation (LSDA). The preliminary results show thatproposed framework can achieve 0.5 mAP (mean Average Precision)within 30 min of response-time for unseen concepts. We expectto improve it further using modied LSDA while applying fastestclassication (MobileNet) and detection (YOLOv3) network, alongwith elimination of requirement of annotated bounding boxes.