Convolutions for Spatial Interaction Modeling

TL;DR

This paper demonstrates that 2D convolutions can effectively model spatial interactions in autonomous vehicle scenarios, offering a faster alternative to graph neural networks with comparable performance, and introduces a new interaction loss for improved accuracy.

Contribution

The paper shows that 2D convolutions can replace GNNs for spatial interaction modeling, reducing latency and introducing a novel interaction loss to enhance modeling quality.

Findings

2D convolutions achieve similar performance to GNNs in interaction modeling.

Convolutions provide lower latency in time-critical systems.

The proposed interaction loss improves modeling accuracy.

Abstract

In many different fields interactions between objects play a critical role in determining their behavior. Graph neural networks (GNNs) have emerged as a powerful tool for modeling interactions, although often at the cost of adding considerable complexity and latency. In this paper, we consider the problem of spatial interaction modeling in the context of predicting the motion of actors around autonomous vehicles, and investigate alternatives to GNNs. We revisit 2D convolutions and show that they can demonstrate comparable performance to graph networks in modeling spatial interactions with lower latency, thus providing an effective and efficient alternative in time-critical systems. Moreover, we propose a novel interaction loss to further improve the interaction modeling of the considered methods.