SparseFusion: Efficient Sparse Multi-Modal Fusion Framework for Long-Range 3D Perception

TL;DR

SarseFusion is introduced, a novel multi-modal fusion framework fully built upon sparse 3D features to facilitate efficient long-range perception and reduces memory footprint and accelerates the inference by about two times compared to dense detectors.

Abstract

Multi-modal 3D object detection has exhibited significant progress in recent years. However, most existing methods can hardly scale to long-range scenarios due to their reliance on dense 3D features, which substantially escalate computational demands and memory usage. In this paper, we introduce SparseFusion, a novel multi-modal fusion framework fully built upon sparse 3D features to facilitate efficient long-range perception. The core of our method is the Sparse View Transformer module, which selectively lifts regions of interest in 2D image space into the unified 3D space. The proposed module introduces sparsity from both semantic and geometric aspects which only fill grids that foreground objects potentially reside in. Comprehensive experiments have verified the efficiency and effectiveness of our framework in long-range 3D perception. Remarkably, on the long-range Argoverse2 dataset, SparseFusion reduces memory footprint and accelerates the inference by about two times compared to dense detectors. It also achieves state-of-the-art performance with mAP of 41.2% and CDS of 32.1%. The versatility of SparseFusion is also validated in the temporal object detection task and 3D lane detection task. Codes will be released upon acceptance.

Figures (7)

• Figure 1: Visualization of sparse features generated from Lift-Splat-Shot (LSS) philion2020lift and our Sparse View Transformer. The pink area denotes non-empty voxels predicted from images. We filter most of the background in the scene, which consumes less memory and computation resources in long-range perception.
• Figure 2: SparseFusion comprises four key modules: the Feature Extractor, Sparse View Transformer, Sparse Encoder, and Detection Head. Initially, the LiDAR point cloud and camera images undergo separate encoders for feature extraction. In the case of camera image features, our Sparse view transformer selectively filters out most of the background information and discards depth values with low confidence, resulting in the generation of sparse 3D features. Subsequently, we fuse features from different modalities and employ a fully sparse encoder for feature extraction. Finally, we apply the detection head to predict the results.
• Figure 3: Different view transformer module. In (a), depth distribution is predicted for each pixel, and all frustums are projected to BEV. In (b), the background of the image is first filtered out, then the top-K values corresponding to the depth distribution of pixels in the foreground region of the image are retained. This significantly reduces the sparsity of the BEV features generated.
• Figure 4: Comparison of the latency and memory cost between our method and BEVFusion.
• Figure 5: The BEV sparsity of SparseFusion under different ranges.