Article Preview
TopIntroduction
With the technological evolution of digital acquisition and storage technologies, millions of images and video sequences are captured every day and shared in online services such as Facebook, Flickr, and Picasa. As keyword-based indexing is very time consuming and inefficient due to linguistic and semantic ambiguities, content-based image and video retrieval systems have been proposed (Vajda, Dufaux, Minh, & Ebrahimi, 2009), which search and retrieve documents based on visual features. Within such systems, a query document is compared to all the documents in the database by making use of content-based features extracted from it. However, since the features are extracted from images which contain two-dimensional projections of three-dimensional scenes, the features may change significantly depending on the view point. Thus, systems often fail to retrieve relevant content in response to some queries.
In general, content-based image retrieval can utilize different representations for describing the image content, including global descriptors, feature points, or regions. Recently, interest has turned towards higher-level representations such as objects. Given a query image containing an object, an image retrieval system can perform two tasks: object recognition or object duplicate detection. Object recognition aims at finding all the instances of a certain object class (such as cars, or shoes), while object duplicate detection represents a more specific task of finding only a particular sample of that object class (such as “red Citroen C3 car” or “white Converse sneakers”). Figure 1 illustrates the relationship between object duplicate detection and object recognition problems. Therefore, within a complete system object recognition is usually applied first to detect a relevant class of objects (e.g., faces, cars) and then object duplicate detection is used to find a specific instance of that object class. Our object duplicate detection system is able to fulfill both tasks together.
Figure 1. Illustration of relationship between object recognition and object duplicate detection. While the former groups objects into different classes such as cars and shoes, the latter distinguishes between specific shoes or cars
In this paper, we are focusing on the object duplicate detection task. The general goal is to detect the presence of a target object in a set of images based on an object model created from a small set of training images. Duplicate objects may vary in their perspective, have different sizes, or be modified versions of the original object after minor manipulations, which do not change their identity. Therefore, object duplicate detection should be robust to changes in position, size, view, illumination, and partial occlusions.
A large number of applications can benefit from object duplicate detection. For example, in the popular photo sharing websites, untagged images can be automatically annotated based on the detection of the same objects from a smaller set of images with associated tags. Also, object duplicate detection may be used to search a specific object in a large collection, such as a suspect car in a video surveillance database. Moreover, when a user takes a picture of an object with his/her mobile phone, additional information about the object can be retrieved from the web, such as the price of a product, or the name and location of a monument.