| Summary: | Corners are stable local image features that have played a crucial role in various computer vision and image processing tasks. Existing corner detectors generally assume that the distance between every two adjacent pixels is a constant during curvature estimation. However, this assumption is actually invalid, and thus those pixel-based measures of discrete curvature developed and exploited in the existing corner detectors may suffer instability under rotation transformations. To address this fundamental problem, a novel curvature measure is proposed in this paper, which exploits the median of a subpixelized triangle located at the current point to estimate its discrete curvature. It is shown that our proposed curvature measure is invariant under rotation transformations. Based on this novel curvature measure, a new corner detector is further developed. Extensive experimental results show that our proposed corner detector can deliver superior performance over the existing state-of-the-art methods.
|
|---|
