This paper presents an unsupervised object extraction system that extracts a single object from natural scenes without relying on color information. The contour information and texture information are analyzed through separate oriented-edge-based processing channels and then combined to complement each other. Contour candidates are extracted from multiresolution edge maps, whereas the local texture information is compactly represented by an oriented-edge-based feature vector and then analyzed by K-means clustering. The object region is determined by merging the results of two separate analysis channels based on the simple assumption that the object is located centrally in the scene. As a result, the object region has been successfully extracted from the scene with a well-defined single boundary line. Both subjective and objective evaluations were carried out and it is shown that the proposed algorithm handles the challenges of complex background well, using only gray-scale images.

A great challenge in real-world applications of action recognition is the lack of sufficient label information because of variance in the recording viewpoint and differences between individuals. A system that can adapt itself according to these variances is required for practical use. We present a generic method for extracting view-invariant features from skeleton joints. These view-invariant features are further refined using a stacked, compact autoencoder. To model the challenge of real-world applications, two unexampled test settings (NewView and NewPerson) are used to evaluate the proposed method. Experimental results with these test settings demonstrate the effectiveness of our method.