String Features: Geodesic Sweeping Detection and Quasi-invariant Time-Series Description

We propose novel features, denoted as string features, which are represented by curves in the image plane. The string features take advantage of its locality at individual points of the curve and of its global aspect when considering the whole curve. The contributions of this paper are: (1) a feature detection procedure which produces a saliency measure by applying a novel technique, named geodesic sweeping, inspired by spatial attention and eye movement control, (2) the description of string features as a set of time-series based on quasi-invariant geometric measures, and (3) a matching algorithm for string features which allows partial matching independently for each time-series in the descriptor. The quantitative performance of the feature detection step is measured with regards to precision, compactness, and repeatability. The repeatability rate reaches 70% with only 3% of the pixels being detected. The string feature matching procedure is tested with a set of 80 synthetic 2D curves. The experimental results show an average ratio of 72.4% in correct matching.