Viewpoint Invariant Object Detector

TL;DR

This paper introduces a viewpoint invariant object detector that enhances robustness to severe viewpoint changes, improves computational efficiency, and demonstrates practical applications in image and video search.

Contribution

The proposed model offers increased robustness to viewpoint variations and addresses computational bottlenecks, improving speed and space efficiency without sacrificing accuracy.

Findings

Enhanced detection performance under <60° viewpoint changes

Significant improvements in model speed and space efficiency

Successful deployment in content-based image and video search

Abstract

Object Detection is the task of identifying the existence of an object class instance and locating it within an image. Difficulties in handling high intra-class variations constitute major obstacles to achieving high performance on standard benchmark datasets (scale, viewpoint, lighting conditions and orientation variations provide good examples). Suggested model aims at providing more robustness to detecting objects suffering severe distortion due to < 60{\deg} viewpoint changes. In addition, several model computational bottlenecks have been resolved leading to a significant increase in the model performance (speed and space) without compromising the resulting accuracy. Finally, we produced two illustrative applications showing the potential of the object detection technology being deployed in real life applications; namely content-based image search and content-based video search.