Scholarlink

<p>Perception is essential for autonomous driving (AD), where accurate and efficient 3D scene understanding from multi-camera inputs remains a core challenge. 3D object detection often struggles with fine-grained scene reconstruction and rare object recognition. In contrast, 3D occupancy prediction offers dense voxel-based representations for safety-critical tasks, yet most existing occupancy prediction methods rely on heavy 3D voxel-based processing, leading to significant computational overhead. Incorporating object detection tasks can provide strong spatial priors that enhance the accuracy and structural consistency of occupancy prediction. Although some recent works leverage detection tasks to aid occupancy prediction, these approaches generally involve multiple network branches or sophisticated post-processing. We propose LSOcc, a lightweight 3D occupancy prediction framework that leverages object detection as a spatial prior to enhance occupancy prediction with a pure vision modality. To enhance occupancy prediction, we introduce a Specific-target Focus Pre-training strategy that performs semantic segmentation on key detection classes within the image backbone. Furthermore, we design an Adaptive Combined Head that fuses detection-aware BEV features into the occupancy head, enhancing occupancy perception-particularly for critical object categories. Experiments on the Occ3D-nuScenes benchmark demonstrate that LSOcc achieves the highest (37.16 mIoU) occupancy performance with only a ResNet-50 backbone and a 256<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times \)</EquationSource> </InlineEquation>704 input resolution, underscoring its accuracy and deployment efficiency for real-world AD applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

LS-Occ:light specific-target-focus vision-based 3D occupancy prediction with adaptive combined head

  • 期刊论文

  • 发表日期：

    2026年2月8日

查看全文
  • Shuo Li,
  • Xu Li,
  • Zihang Wang

摘要

Perception is essential for autonomous driving (AD), where accurate and efficient 3D scene understanding from multi-camera inputs remains a core challenge. 3D object detection often struggles with fine-grained scene reconstruction and rare object recognition. In contrast, 3D occupancy prediction offers dense voxel-based representations for safety-critical tasks, yet most existing occupancy prediction methods rely on heavy 3D voxel-based processing, leading to significant computational overhead. Incorporating object detection tasks can provide strong spatial priors that enhance the accuracy and structural consistency of occupancy prediction. Although some recent works leverage detection tasks to aid occupancy prediction, these approaches generally involve multiple network branches or sophisticated post-processing. We propose LSOcc, a lightweight 3D occupancy prediction framework that leverages object detection as a spatial prior to enhance occupancy prediction with a pure vision modality. To enhance occupancy prediction, we introduce a Specific-target Focus Pre-training strategy that performs semantic segmentation on key detection classes within the image backbone. Furthermore, we design an Adaptive Combined Head that fuses detection-aware BEV features into the occupancy head, enhancing occupancy perception-particularly for critical object categories. Experiments on the Occ3D-nuScenes benchmark demonstrate that LSOcc achieves the highest (37.16 mIoU) occupancy performance with only a ResNet-50 backbone and a 256 \(\times \) 704 input resolution, underscoring its accuracy and deployment efficiency for real-world AD applications.