RfD-Net: Point Scene Understanding by Semantic Instance Reconstruction

In this section, we provide all the parameters, layer specifications and weights used in loss functions. We uniformly denote the fully-connected layers by MLP [l 1 , ..., l d ], where l i is the number of neurons in the i-th layer. A.1. 3D Detector In section 3.1, we predict object proposals from N input points with VoteNet [9] as the backbone. It produces N p proposals with D p -dim features (i.e. proposal features F p ∈ R Np×Dp in our paper), from which we regress the D b -dim box parameters with MLP [128, 128, 69] (N =80K, N p =256, D p =128, D b =69). As in [9], the 69-dim box parameters are encoded by center c ∈ R 3 , scale s 3 ∈ R 3 , heading angle θ ∈ R, semantic label l, and objectness score s obj . s obj is a probability value indicating whether the proposal is close to (<0.3 meter, positive) or far from (>0.6 meter, negative) any ground-truth object center.