Problem Introduction and Background
Human brain is inherently capable of performing complex object recognition tasks. We are capable of using a variety of visual cues to deduce the the class or identity of an object. For example, when we look at a car, we use it’s structural information to deduce that it’s a car and not a bus or a truck. The appearance of the car and it’s properties such as length, height, ground clearance, wheelbase and track etcetera allow us to recognize it as a car. Depending on the level of sophistication of our training in recognizing the cars we can further deduce the make and model of the car, year of manufacturing etc. Identifying object class in a given image, however, is usually a nontrivial task for a computer. To identify the instances of an object in a given image, a computer must have the representation or description of the object class, and a time-efficient way of searching for the object instances with matching descriptions in the given image.
Object class recognition by a computer is usually organized as follows: 1. An object description or “quantified definition” of object class is obtained – these are the visual cues for the computer algorithm, 2. Computer algorithm is trained to learn the description of the objects, 3. Algorithm is tested to recognize the object class instances in the test image. Next section describes the planned approach to accomplish these three steps.
Method
The complete technique that is planned to be used can be described as a two-step process:
-
1.Object Description or Representation : Using the Objectness score as defined by Ferrari et al., [1] it’s possible to find multiple windows (sub-images) within an image that have higher chances of containing an object as compared to other sub-images of the same image. The idea is to let the algorithm compute the object description by analyzing the images which contain instance(s) of the particular object class. To facilitate this process, windows of an image that have high objectness score will be input to a Multiple Instance Learning (MIL) algorithm [2], [3]. Thus, for each image containing the instance(s) of the object class, algorithm will analyze several sub-images. By analyzing images from the training set, the MIL algorithm will develop an abstraction of the object class. In essence, this step combines step 1 and 2 described in the previous paragraph.
-
2.To find the instances of a given object within a test image, the algorithm will be provided with the sub-images of the test image that have high chances of containing a generic object. This will help in speeding up the object class recognition process.
Dataset
PASCAL VOC 2007 dataset [4] will be used for this work. This dataset consists of over 5000 images, which contain multiple instances of 20 object classes, which can be majorly classified as person, animal, vehicle or indoor. This dataset has been used by several research groups that are working on object recognition. Thus, it will serve as a standard for comparison of different algorithms.
Milestone Goal
By May 8th ’12, I intend to complete a comprehensive literature review and develop the software to implement step 1 described in Method.
References
[1] Bogdan Alexe, Thomas Deselaers, and Vittorio Ferrari, “Measuring the objectness of image windows”, IEEE Transactions on Pattern Analysis and Machine Intelligence. (Early print online).
[2] Matthew Blaschko and Thomas Hoffman, “Conformal Multi-Instance Kernels”, NIPS 2006 Workshop on Learning to Compare Examples, 2006.
[3] Jun Yang, “Review of Multi-Instance Learning Algorithms”, Web: http://www.cs.cmu.edu/~juny/MILL/mil_review.pdf
[4] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn and A. Zisserman, "The PASCAL Visual Object Classes Challenge 2007 Results", Web: http://www.pascal-network.org/challenges/VOC/voc2007/workshop/index.html