Abstract
Balancing lightweight model design with detection accuracy remains a major challenge in 3D object detection. To address this issue, we propose a novel Frustum DeepScaleNet (F-DSNet) based on the Frustum ConvNet framework. F-DSNet integrates a Res2Net-style hierarchical residual multi-scale design by embedding hierarchical residual connections into the convolutional layers and replacing each single convolution branch with parallel multi-scale pathways. This design enhances context aggregation with an enlarged effective receptive field while reducing parameter redundancy in the feature extraction stage, enabling architectural-level model compression. Additionally, we propose a bounding-box quality estimation module. This module jointly models classification confidence and center localization error to boost contributions from high-quality bounding boxes, suppress adverse effects from low-quality boxes, and improve both classification and localization performance. Experimental results on the SUN RGB-D and KITTI datasets show that the 4-layer F-DSNet reduces the number of parameters by 32.23% and FLOPs by 0.73%, while achieving a 0.4% improvement in mAP over the baseline F-ConvNet. Furthermore, its 3-layer variant achieves a 61.91% reduction in parameters and a 4.41% reduction in FLOPs, while maintaining comparable detection accuracy.
Funding Information
  • National Natural Science Foundation of China (62276042)
  • The Basic Scientific Research Program of the Department of Education of Liaoning Province (LJKMZ20220838)

This publication has 16 references indexed in Scilit: