SCAFusion: A Multimodal 3D Detection Framework for Small Object Detection in Lunar Surface Exploration

TL;DR

SCAFusion addresses the critical need for reliable small-object detection in lunar surface perception by extending BEVFusion with four components: a parameter-efficient Cognitive Adapter Mona, a Contrastive Alignment Module for RGB-depth consistency, a Section-aware Coordinate Attention mechanism tailored for small targets, and a Camera Auxiliary Training Branch to boost visual representation during training. The approach yields significant gains on lunar simulation data (Isaac Sim) and competitive improvements on the nuScenes benchmark, notably enhancing small-object detection while maintaining low inference overhead. Key contributions include explicit small-object-focused attention, cross-modal alignment, and training-time camera supervision that collectively strengthen multimodal fusion under off-world constraints. The results demonstrate practical relevance for autonomous lunar missions and establish a pathway toward robust, generalizable multimodal perception in challenging, resource-constrained environments.

Abstract

Reliable and precise detection of small and irregular objects, such as meteor fragments and rocks, is critical for autonomous navigation and operation in lunar surface exploration. Existing multimodal 3D perception methods designed for terrestrial autonomous driving often underperform in off world environments due to poor feature alignment, limited multimodal synergy, and weak small object detection. This paper presents SCAFusion, a multimodal 3D object detection model tailored for lunar robotic missions. Built upon the BEVFusion framework, SCAFusion integrates a Cognitive Adapter for efficient camera backbone tuning, a Contrastive Alignment Module to enhance camera LiDAR feature consistency, a Camera Auxiliary Training Branch to strengthen visual representation, and most importantly, a Section aware Coordinate Attention mechanism explicitly designed to boost the detection performance of small, irregular targets. With negligible increase in parameters and computation, our model achieves 69.7% mAP and 72.1% NDS on the nuScenes validation set, improving the baseline by 5.0% and 2.7%, respectively. In simulated lunar environments built on Isaac Sim, SCAFusion achieves 90.93% mAP, outperforming the baseline by 11.5%, with notable gains in detecting small meteor like obstacles.

Figures (10)

• Figure 1: Overall pipeline of SCAFusion. Firstly, the model extracts multi-modal features through specific backbones. Then image features are transformed into image-BEV features after enhancing semantic consistency through Contrastive Alignment Module. Concurrently, image-BEV features fuse with lidar-BEV features and improve the detection capability of small target objects through the Section-aware Coordinate Attention module. The Cognitive Adapter in the camera backbone freezes the pre-training weights of the original camera backbone during backpropagation and adjust fewer parameters to achieve better performance. Only during the training process, image-BEV features are processed through Camera Auxiliary Training Branch(CATB) to fully tap the potential of camera information in multimodal fusion process.
• Figure 2: Illustration of Cognitive-Adapter-based backbone.
• Figure 3: Illustration of Contrastive Alignment Module. RGB features are aligned with depth features through the alignment loss.
• Figure 4: Illustration of Section-aware Coordinate Attention Module.
• Figure 5: Illustration of Camera Auxiliary Training Branch.