Boosting the Performance of Object Tracking with a Half-Precision Particle Filter on GPU

TL;DR

This paper demonstrates that using half-precision on GPU for particle filters significantly boosts performance with minimal accuracy loss, benefiting real-time object detection tasks.

Contribution

The paper introduces optimized half-precision particle filter algorithms on CUDA, achieving substantial speedups over traditional precision levels while maintaining acceptable accuracy.

Findings

Performance improved by 1.5-2x over single-precision

Performance improved by 2.5-4.6x over double-precision

Minor accuracy loss with half-precision implementation

Abstract

High-performance GPU-accelerated particle filter methods are critical for object detection applications, ranging from autonomous driving, robot localization, to time-series prediction. In this work, we investigate the design, development and optimization of particle-filter using half-precision on CUDA cores and compare their performance and accuracy with single- and double-precision baselines on Nvidia V100, A100, A40 and T4 GPUs. To mitigate numerical instability and precision losses, we introduce algorithmic changes in the particle filters. Using half-precision leads to a performance improvement of 1.5-2x and 2.5-4.6x with respect to single- and double-precision baselines respectively, at the cost of a relatively small loss of accuracy.