A convolutional neural networks oriented approach for voxel-based 3D object classification

In our work, 3D objects classification has been dealt with convolutional neural networks which is a common paradigm recently in image recognition. In the first phase of experiments, 3D models in ModelNet10 and ModelNet40 data sets were voxelized and scaled with certain parameters. Classical CNN and 3D Dense CNN architectures were designed for training the pre-processed data. In addition, the two trained CNNs were ensembled and the results of them were observed. A success rate of 95.37% achieved on ModelNet10 by using 3D dense CNN, a success rate of 91.24% achieved with ensemble of two CNNs on ModelNet40.