Optimizing GNN-based multiple object tracking on a graphcore IPU

Abstract

This paper improves multi-object tracking (MOT) efficiency using Graphcore's IPUs with GNNs. GNNs are crucial in real-time applications like autonomous driving and robotics for modeling complex object interactions, yet their computational demands, especially in key message passing operations, hinder performance. We discuss adapting a PyTorch model to TensorFlow for IPU execution and compare IPU and GPU performance. Baseline metrics such as average training and inference time per epoch are assessed, providing insights into each platform's strengths and limitations. We then focus on optimizing message passing operations for GNN efficiency on IPUs, evaluating the effects of these optimizations and adjustments to IPU-specific configurations.