SuBloNet: Sparse Super Block Networks for Large Scale Volumetric Fusion

Abstract

Training and inference of convolutional neural networks (CNNs) on truncated signed distance fields (TSDFs) is a challenging task. Large parts of the scene are usually empty, which makes dense implementations inefficient in terms of memory consumption and compute throughput. However, due to the truncation distance, non-zero values are grouped around the surface creating small dense blocks inside the large empty space. We show that this structure can be exploited by storing the TSDF in a block sparse tensor and then decomposing it into rectilinear super blocks. A super block is a dense 3d cuboid of variable size and can be processed by conventional CNNs. We analyze the rectilinear decomposition and present a formulation for computing the bandwidth-optimal solution given a specific network architecture. However, this solution is NP-complete, therefore we also a present a heuristic approach for fast training and inference tasks. We verify the effectiveness of SuBloNet and report a speedup of 4x towards dense implementations and 1.7x towards state-of-the-art sparse implementations. Using the super block architecture, we show that recurrent volumetric fusion is now possible on large scale scenes. Such a systems is able to reconstruct high-quality surfaces from few noisy depth images.