MSFASA-3DNet: Multi-Scale Feature Aggregation and Spatial Attention for 3-D Object Detection

Perception is the most important component in autonomous vehicles and robots. LiDAR-based 3D object detection (3D-OD) provides accurate spatial information for precise object localization. Traditional image-based 3D-OD methods struggle with depth estimation, making them less reliable for 3D scene understanding. The existing LiDAR-based model uses single or two-stage convolutional networks to perform 3D-OD tasks. The existing models are computationally expensive with complex backbones. We propose an enhanced backbone with multi-scale feature extraction and spatial attention to improve LiDAR-based 3D-OD. Our backbone is computationally effective as we have reduced almost 40% of the parameters of existing backbones and achieved competitive accuracy compared to existing models. Multi-scale feature extraction captures object details at different resolutions to provide fine and coarse features. Spatial attention concentrates on important regions and suppresses background noise to generate more accurate spatial detections. Our backbone integrates multi-scale feature aggregation and spatial attention mechanisms to improve detection accuracy. Our model uses the standard single-stage and grid-based 3D-OD pipeline with our proposed backbone. We conducted extensive ablation experiments to validate our proposed backbone’s efficiency. Our model outperforms existing models by being lightweight, which makes it highly suitable for real-time applications. We evaluated our model using the KITTI benchmark. Our results are featured on the official leaderboard of KITTI.