From Feature Detection in Truncated Signed Distance Fields to Sparse Stable Scene Graphs

With the increased availability of GPUs and multicore CPUs, volumetric map representations are an increasingly viable option for robotic applications. A particularly important representation is the truncated signed distance field (TSDF) that is at the core of recent advances in dense 3-D mapping. However, there is relatively little literature exploring the characteristics of 3-D feature detection in volumetric representations. In this letter, we evaluate the performance of features extracted directly from a 3-D TSDF representation. We compare the repeatability of Integral invariant features, specifically designed for volumetric images, to the 3-D extensions of Harris and Shi & Tomasi corners. We also study the impact of different methods for obtaining gradients for their computation. We motivate our study with an example application for building sparse stable scene graphs, and present an efficient GPU-parallel algorithm to obtain the graphs, made possible by the combination of TSDF and 3-D feature points. Our findings show that while the 3-D extensions of 2-D corner-detection perform as expected, integral invariants have shortcomings when applied to discrete TSDFs. We conclude with a discussion of the cause for these points of failure that sheds light on possible mitigation strategies.