Bias-Induced Point Auto-Encoder and Comparative Analysis of Point Encoder-Decoder Combinations

As a combination of a point encoder and a decoder, a point auto-encoder (AE) facilitates the reconstruction and segmentation of the point cloud representation of 3D objects. In this paper, first, we present a novel approach to point decoding, referred to here as bias-induced point decoding, which learns pointwise decoding of the same global feature by the layer-wise control of mapping biases while sharing mapping weights, instead of pointwise differentiating the input by concatenating local features or grids to the global feature as is done conventionally. Then, in-depth comparative analyses of 12 possible encoder-decoder combinations were conducted to identify the strengths and weaknesses in terms of accuracy and robustness in reconstruction and segmentation. We verified that a bias-induced point AE could be a viable alternative to conventional point AEs with its own strength and weakness. Finally, the effectiveness of the proposed bias-induced point AE was demonstrated based on its application to the automatic transformation from partial-to-full 3D point cloud representations.