Improvements to the TCVD method to segment hand-drawn sketches

Abstract

Tangent and Corner Vertices Detection (TCVD) is a method to detect corner vertices and tangent points in sketches using parametric cubic curves approximation, which is capable to detect corners with a high accuracy and a very low false positive rate, and also to detect tangent points far above other methods in literature. In this article, we present several improvements to TCVD method in order to establish mathematical conditions to detect corners and make the obtaining of curves independent from the scale, what increases the success ratio in transitions between lines and curves. The new conditions for obtaining corners use the radius as the inverse of the curvature, and the second derivative of the curvature. For the detection of curves, a new descriptor is presented, avoiding the parameters dependent of scale used in TCVD method. In order to obtain the performance of the implemented improvements, several tests have been carried out using a dataset which contains sketches more complex than those used for validation of TCVD algorithm (sketches with more curves and tangent points and sketches of different sizes). For corners detection, the accuracy obtained was pretty similar to that obtained with the previous TCVD, however, for curves and tangent points detection the accuracy increases significantly.