During the 3D reconstruction of a unknown object, carried out by ground or air vehicles, the Next Best View (NBV) is computed to decide the next sensor position. This work is motivated by the fact that new Deep Learning approaches are incorporating volumetric representations of the space and had have a big success in pattern recognition tasks. Therefore, we present a labeled dataset for the NBV problem. The dataset contains examples of partial models and their corresponding NBV class. A probabilistic grid is used to represent the partial model of the reconstruction. To build the dataset, we pre-process the RGB-D images used in the SIXD challenge and we have converted them into point clouds and perceptions. Such perceptions are used to reconstruct several times each object. The aim of this work is to provide a dataset which will be useful to train an artificial neural network to solve the NBV problem. The dataset is available for the community.
This dataset was proposed in Mendoza’s master thesis (only in spanish), the research paper is currently under review.
It can be downloaded from:
https://www.kaggle.com/miguelmg/nbv-dataset
NBV-net implementation in TensorFlow:
https://github.com/MiguelMendozaG/Programas_NBV