Object Recognition Under Difficult Conditions Based on Superpixel

In computer vision the task of object recognition is to recognize a certain object in an provided image. For this task a description about the object of interest is learned from images. This process often involves the use of local features like SIFT [1] which can be extracted reliably from images, even if the resolution or the lighting conditions change. The drawback of such local features are that they may lose discriminative power to distinguish between similar objects or the object from the background. We proposed a system that utilizes superpixel [2]. The system is not only working on local features rather it packs the features belonging to one such superpixel and reject the whole area, if it is ambiguous. A superpixel is ambiguous if the system can not name an object with a high confidence to which object this superpixel could belong. By doing so we have achieved an improvement of over 10% on a difficult dataset.