A Two Fold Approach for Object Recognition with Bag of Visual Words using Artificial Neural Network

Abstract

Advancements in Computer vision techniques have assisted the Artificial intelligence and Autonomous robotics community. Its applications involve driverless cars, Visual Odometry, Object detection and localization and 3D Reconstruction. Object recognition and localization have also been a hot topic for researchers in recent years. Object detection is a very important part in driverless cars too, i.e., it detects objects such as humans and other vehicles. Proposed is a technique to recognize and localize the object in the image effectively. This Proposed technique first reduces the problem from multi-class, multi-label classification to a single label, multi-class problem by only taking a bag of visual word (BoVW) model of a single bounding box at a time. Moreover, this also generates ‘none’ class edge boxes and suppresses these ‘none’ objects based on its BoVW model. PASCAL VOC 2012 and 2007 datasets were used for training and testing purposes. Both qualitative and quantitative analysis was carried out. It is established that the proposed approach gives appreciable results while reducing the complexity of the problem.