A Novel Approach to Object Detection: Object Search

Abstract

Most object detection algorithms attempt to detect all objects present in an image and accordingly classify them. While that approach is useful for various domains and applications, there are also many cases where we would only want to search for a particular object in a given image. For such cases, there is potential to optimize the search by focusing on the object we are looking for and ignoring the rest of the information in the image to the maximum possible extent, thereby greatly improving the computation speed. In this light, I have developed a model which can search for an object given in an image (the object image) in another image where the object mayor may not be present (the target image). The design takes inspiration from Siamese Neural Networks and techniques applied in other object detection algorithms and combines them with a novel technique and loss. I have trained and tested the model using images from the COCO dataset. It has shown improvement in computation speed compared to other state-of-the-art models for the desired task, along with appreciable accuracy.