Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DCDet: Dynamic Cross-based 3D Object Detector

Shuai Liu, Boyang Li, Zhiyu Fang, Kai Huang

TL;DR

DCDet tackles two core challenges in 3D object detection: imbalanced positive sampling across object scales and rotation-aware regression. It introduces Dynamic Cross Label Assignment (DCLA), which samples positives from a cross-shaped region around each ground-truth center using a cost-based top-$k$ selection and a scalable radius $r$, and Rotation-Weighted IoU (RWIoU), which integrates rotation and direction into the IoU to drive regression. Together, these yield a unified framework that achieves state-of-the-art or competitive performance on Waymo Open and KITTI, with notable data efficiency and improved performance for small objects. The approach highlights the importance of task-aligned label assignment and rotation-aware loss design for robust 3D detection in real-world scenarios.

Abstract

Recently, significant progress has been made in the research of 3D object detection. However, most prior studies have focused on the utilization of center-based or anchor-based label assignment schemes. Alternative label assignment strategies remain unexplored in 3D object detection. We find that the center-based label assignment often fails to generate sufficient positive samples for training, while the anchor-based label assignment tends to encounter an imbalanced issue when handling objects of varying scales. To solve these issues, we introduce a dynamic cross label assignment (DCLA) scheme, which dynamically assigns positive samples for each object from a cross-shaped region, thus providing sufficient and balanced positive samples for training. Furthermore, to address the challenge of accurately regressing objects with varying scales, we put forth a rotation-weighted Intersection over Union (RWIoU) metric to replace the widely used L1 metric in regression loss. Extensive experiments demonstrate the generality and effectiveness of our DCLA and RWIoU-based regression loss. The Code will be available at https://github.com/Say2L/DCDet.git.

DCDet: Dynamic Cross-based 3D Object Detector

TL;DR

DCDet tackles two core challenges in 3D object detection: imbalanced positive sampling across object scales and rotation-aware regression. It introduces Dynamic Cross Label Assignment (DCLA), which samples positives from a cross-shaped region around each ground-truth center using a cost-based top- selection and a scalable radius , and Rotation-Weighted IoU (RWIoU), which integrates rotation and direction into the IoU to drive regression. Together, these yield a unified framework that achieves state-of-the-art or competitive performance on Waymo Open and KITTI, with notable data efficiency and improved performance for small objects. The approach highlights the importance of task-aligned label assignment and rotation-aware loss design for robust 3D detection in real-world scenarios.

Abstract

Recently, significant progress has been made in the research of 3D object detection. However, most prior studies have focused on the utilization of center-based or anchor-based label assignment schemes. Alternative label assignment strategies remain unexplored in 3D object detection. We find that the center-based label assignment often fails to generate sufficient positive samples for training, while the anchor-based label assignment tends to encounter an imbalanced issue when handling objects of varying scales. To solve these issues, we introduce a dynamic cross label assignment (DCLA) scheme, which dynamically assigns positive samples for each object from a cross-shaped region, thus providing sufficient and balanced positive samples for training. Furthermore, to address the challenge of accurately regressing objects with varying scales, we put forth a rotation-weighted Intersection over Union (RWIoU) metric to replace the widely used L1 metric in regression loss. Extensive experiments demonstrate the generality and effectiveness of our DCLA and RWIoU-based regression loss. The Code will be available at https://github.com/Say2L/DCDet.git.
Paper Structure (22 sections, 13 equations, 3 figures, 9 tables)

This paper contains 22 sections, 13 equations, 3 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. 3D Object Detection
  4. Label Assignment
  5. IoU-based Loss
  6. Methodology
  7. Dynamic Cross Label Assignment
  8. Rotation-Weighted IoU
  9. Loss Function
  10. Experiments
  11. Implementation Setup
  12. Data Preprocessing
  13. Training Details
  14. Comparison with State-of-the-Art Methods
  15. Effect on Different Backbone Networks
  16. ...and 7 more sections

Figures (3)

  • Figure 1: Cross-shaped region for different grid cell sizes.
  • Figure 2: The overall framework of our DCDet. The dynamic cross label assignment scheme is only used in the training phase.
  • Figure 3: The calculation process of RWIoU.