Abstract:
Analysis of point clouds through deep convolutional neural networks is an active area
of research due to their massive real-world applications including autonomous driving,
indoor navigation, robotics, virtual/augmented reality, unmanned aerial vehicles and
drones technology. However, to capture fine grained geometric and semantic properties
for the underlying recognition task with raw point cloud is exceedingly challenging due to
their irregular and unordered nature, sparsity and lack of implicit neighborhood. In this
paper, we have introduced a deep, hierarchical, 3d point based architecture to address
the highly challenging problem of object classification and part segmentation using raw
point cloud. The proposed architecture consists of multiple layers of Sampling, Annular
convolution and Pooling, cascaded together in accordance with the principle of deep
residual learning. In the skip connections of our deep residual design, we propose to
use a combination of linear Projection shortcut and nonlinear Relu group normalization
shortcut with batch normalization, to improve both the optimization landscape and
representational power. Our network achieves on par or even better than state of the
art results on synthetic and real-world benchmark datasets of object classification i.e.
MODELNET40 and ScanObjectNN and part segmentation i.e. ShapeNet-part.
