Object Detection with Real-Time Context Adaptation

Abstract: Visual object detection based on Convolutional Neural Networks (CNNs) isquickly becoming an essential component of many Augmented Reality (AR) applications.Despite of the impressive performance, these state-of-the-art systemshave one obvious weakness: they do not have capability to learn at run-time,from the visual appearance of the target object in a particular test video.To address this issue, a real-time content-adaptive algorithm that learns fromthe data on-the- y is built. The solution is based on a novel algorithm: RegularizedOnline-Sequential Extreme Learning Machines (OS-ELM). The proposedapproach outperforms the commonly used Incremental Support Vector Machines(SVM) architecture, both in terms of complexity and accuracy. Using the proposedOS-ELM as a post-processor to Faster R-CNN with VGG model improvesthe detection accuracy of this popular system.