BINK: Biological binary keypoint descriptor

Learning robust keypoint descriptors has become an active research area in the past decade. Matching local features is not only important for computational applications, but may also play an important role in early biological vision for disparity and motion processing. Although there were already some floating-point descriptors like SIFT and SURF that can yield high matching rates, the need for better and faster descriptors for real-time applications and embedded devices with low computational power led to the development of binary descriptors, which are usually much faster to compute and to match. Most of these descriptors are based on purely computational methods. The few descriptors that take some inspiration from biological systems are still lagging behind in terms of performance. In this paper, we propose a new biologically inspired binary keypoint descriptor: BINK. Built on responses of cortical V1 cells, it significantly outperforms the other biologically inspired descriptors. The new descriptor can be easily integrated with a V1-based keypoint detector that we previously developed for real-time applications.

Multi-scale keypoints in V1 and beyond: object segregation, scale selection, saliency maps and face detection

End-stopped cells in cortical area V1, which combine outputs of complex cells tuned to different orientations, serve to detect line and edge crossings, singularities and points with large curvature. These cells can be used to construct retinotopic keypoint maps at different spatial scales (level-of-detail). The importance of the multi-scale keypoint representation is studied in this paper. It is shown that this representation provides very important information for object recognition and face detection. Different grouping operators can be used for object segregation and automatic scale selection. Saliency maps for focus-of-attention can be constructed. Such maps can be employed for face detection by grouping facial landmarks at eyes, nose and mouth. Although a face detector can be based on processing within area V1, it is argued that such an operator must be embedded into dorsal and ventral data streams, to and from higher cortical areas, for obtaining translation-, rotation- and scale-invariant detection.