3D shape retrieval by Poisson histogram

Defining a compact and robust shape descriptor is the most important issue in 3D shape retrieval. This paper proposes a new shape descriptor called Poisson histogram, which can capture shape structure feature very well. Moreover, it is robust under different geometry processing and has low feature dimension for efficient indexing. Poisson histogram can be defined by the following two steps. Firstly, we use Poisson equation to define a 3D shape signature. Secondly, we derive a histogram-based shape descriptor by accumulating the values of the defined signature in bins. To verify the performance of Poisson histogram, we perform an experimental comparison on McGill database of 3D shapes. Results show that Poisson histogram is better than several existing histogram-based shape descriptors both in retrieving accuracy and retrieving efficiency.

Research highlights
► We proposed a new shape descriptor called Poisson histogram.
► Poisson histogram has a good distinction for 3D shapes with different part structures.
► Poisson histogram is robust under pose variation and some geometry processing operations.
► Poisson histogram is better than several descriptors both in accuracy and efficiency.