MGSC3D: Multi-Scale Gated Sparse Convolutional 3D Object Detection
Xiaoyun Yin, Xiangyue Zhang, Peng Ji, Jingyu Ru, Chengdong Wu · 2025
In recent years, point cloud-based 3D object detection has attracted significant attention due to its wide applications in fields such as autonomous driving, robotic navigation, and augmented reality. Traditional sparse convolution algorithms typically use uniform convolution kernels, which may weaken the ability to capture fine-grained details and large-scale features, leading to issues such as missed detection of small objects or misidentification of large objects. Additionally, the inherent unordered nature of point cloud data results in poor performance in modeling point relationships and understanding the overall geometric structure. To address these issues, we propose the MGSC3D algorithm, which combines multi-scale gated sparse convolution and a point cloud position encoder. We also design a decoupled 3DIoU-based loss function for training. Experimental results show that MGSC3D outperforms existing methods on both the ScanNet and SUN RGB-D datasets.