SparseDet: Towards efficient multi-view 3D object detection via sparse scene representation

Abstract

Efficient and reliable 3D object detection via multi-view cameras is pivotal for improving the safety and facilitating the cost-effective deployment of autonomous driving systems. However, owing to the learning of dense scene representations, existing methods still suffer from high computational costs and excessive noise, affecting the efficiency and accuracy of the inference process. To overcome this challenge, we propose SparseDet, a model that exploits sparse scene representations. Specifically, a sparse sampling module with category-aware and geometry-aware supervision is first introduced to adaptively sample foreground features at both semantic and instance levels. Additionally, to conserve computational resources while retaining context information, we propose a background aggregation module designed to compress extensive background features into a compact set. These strategies can markedly diminish feature volume while preserving essential information to boost computational efficiency without compromising accuracy. Due to the efficient sparse scene representation, our SparseDet achieves leading performance on the widely used nuScenes benchmark. Comprehensive experiments validate that SparseDet surpasses the PETR while reducing the decoder computational complexity by 47% in terms of FLOPs, facilitating a leading inference speed of 35.6 FPS on a single RTX3090 GPU.