More Effective Supervised Learning in Randomized Trees for Feature Recognition

This paper presents a feature recognition method based on randomized trees. We aim to improve the performance of Lepetit's work, whose actual results are very sensitive to large changes of viewpoint due to its limited ability of samples synthesizing and learning. We propose an approach to alleviate its limitation, which simulates the image appearance changes under actual viewpoint changes by applying general projective transformations to the standard image rather than affine ones. Affine transformations are usually used in many state-of-the-arts but they cannot adequately represent the actual relationship between two images with different viewpoints. The result is a more effective way of supervised image sample learning in randomized trees for feature recognition that is robust to large changes of viewpoints.