SIFT-based Segmentation of Multiple Instances of Low-Textured Objects
Paolo Piccinini, Andrea Prati, Rita Cucchiara · International Journal of Computer Theory and Engineering · 2013
This paper proposes a simple yet effective approach for segmenting multiple instances of the same object for a pick-and-place application. The considered objects present several challenges, such as low texture, semi-transparent container, moving parts, and severe occlusions. Real-time constraints must be met, calling for a good trade-off between accuracy and efficiency. For all these reasons, the proposed approach is based on SIFT features and a suitable modification of the 2NN matching procedure to increase the number of available matches. Moreover, in order to reduce false segmentations, ad-hoc algorithms based on overlap detection and color similarity are used.