JetsonHacks

NVIDIA Jetson TK1 scores well in SLAMBench introduction

Researchers in the UK have published a paper “Introducing SLAMBench, a performance and accuracy benchmarking methodology for SLAM” which included the NVIDIA Jetson TK1 in their testing.

SLAMBench is a publicly available software framework which provides a KinectFusion implementation in C++, OpenMP, OpenCL and CUDA. Simultaneous localization and mapping (SLAM) techniques are now crossing into mainstream products via projects like Dyson 360 Eye Robot Vacuum and Project Tango.

From the paper:

To address these issues, we present SLAMBench which is a portable implementation of KinectFusion with verifiable accuracy relative to a known ground truth. SLAMBench allows computer vision researchers to provide and evaluate alternative algorithmic implementations but it also enables domain-specific optimisation, auto-tuning and domain-specific languages (DSLs) in dense SLAM. It takes a step forward improving the way academia and industry evaluate system design and implementation in computer vision. To the best of our knowledge, SLAMBench is the first performance, energy and accuracy benchmark dedicated to 3D scene understanding applications.

One of the things I found interesting in the paper is comparing a TITAN desktop, which achieves a frame rate of 135 FPS versus the Jetson which reaches almost real time performance with 22 FPS! Way impressive when the power profile is taken into consideration. That didn’t quite prepare me for the difference of the algorithms GPU vs CPU implementation on the Jetson. Using the CUDA cores, they scored 22 FPS on the GPU. Using just the ARM cores, they scored 0.8 FPS. Wow.

A good read.