Is visual similarity sufficient for semantic object recognition?

The paper discusses experiments (using exemplary classes of man-made objects) on the-same-class object detection based on the keypoint matching techniques. Two algorithms are used, i.e. building clusters of consistently similar and distributed keypoints, and matching individual points represented by novel descriptors incorporating semi-local geometry of images. It is shown that although detection of near-identically looking objects in random images can be performed reliably, the same is not possible for semantically defined classes of objects (even if we expect a certain level of visual and configurational uniformity within the class). The experiments conducted on PASCAL2007 dataset provide results which are not better than random selection. However, selected experimental results indicate that for certain classes of objects semantics may be significantly correlated with the visual and configurational consistencies.