Effectiveness of the Bag-of-Words approach on the object search problem in 3D domain

Bag-of-Words, object search, large-scale datasets

In this work, we investigate the application of the Bag-of-Words approach for object search task in 3D domain. Image retrieval task solutions, operating on datasets of thousands and millions images, have proved the effectiveness of Bag-of-Words approach. The availability of low cost RGB-D cameras is a rise of large datasets of 3D data similar to image corpuses (e.g. RoboEarth). The results of such an investigation could be useful for many robot scenarios like place recognition from a large dataset of samples of places acquired during the long-term observation of an environment. The first goal of our research presented in this paper is focused on the sensitivity of the Bag-of-Words approach to various parameters (e.g. spacial sampling, surface description etc.) with respect to precision, stability and robustness. The experiments are carry out on two widely-used datasets in object instance identification task in 3D domain.