Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards Cross-Platform Generalization: Domain Adaptive 3D Detection with Augmentation and Pseudo-Labeling

Xiyan Feng, Wenbo Zhang, Lu Zhang, Yunzhi Zhuge, Huchuan Lu, You He

TL;DR

This work tackles cross-platform generalization in LiDAR-based 3D object detection by extending PVRCNN++ with Cross-platform Jitter Alignment (CJA) data augmentation and ST3D self-training. The method replaces the CenterHead with an AnchorHead, uses RoI-grid pooling and VSA for robust proposals, and applies a two-stage training regime to adapt to unlabeled target domains. Ablation shows substantial gains from CJA and ST3D, with the final AnchorHead variant achieving state-of-the-art-like performance in cross-platform settings. The approach demonstrates practical cross-platform deployment potential for autonomous-robot perception across vehicles, drones, and quadrupeds.

Abstract

This technical report represents the award-winning solution to the Cross-platform 3D Object Detection task in the RoboSense2025 Challenge. Our approach is built upon PVRCNN++, an efficient 3D object detection framework that effectively integrates point-based and voxel-based features. On top of this foundation, we improve cross-platform generalization by narrowing domain gaps through tailored data augmentation and a self-training strategy with pseudo-labels. These enhancements enabled our approach to secure the 3rd place in the challenge, achieving a 3D AP of 62.67% for the Car category on the phase-1 target domain, and 58.76% and 49.81% for Car and Pedestrian categories respectively on the phase-2 target domain.

Towards Cross-Platform Generalization: Domain Adaptive 3D Detection with Augmentation and Pseudo-Labeling

TL;DR

This work tackles cross-platform generalization in LiDAR-based 3D object detection by extending PVRCNN++ with Cross-platform Jitter Alignment (CJA) data augmentation and ST3D self-training. The method replaces the CenterHead with an AnchorHead, uses RoI-grid pooling and VSA for robust proposals, and applies a two-stage training regime to adapt to unlabeled target domains. Ablation shows substantial gains from CJA and ST3D, with the final AnchorHead variant achieving state-of-the-art-like performance in cross-platform settings. The approach demonstrates practical cross-platform deployment potential for autonomous-robot perception across vehicles, drones, and quadrupeds.

Abstract

This technical report represents the award-winning solution to the Cross-platform 3D Object Detection task in the RoboSense2025 Challenge. Our approach is built upon PVRCNN++, an efficient 3D object detection framework that effectively integrates point-based and voxel-based features. On top of this foundation, we improve cross-platform generalization by narrowing domain gaps through tailored data augmentation and a self-training strategy with pseudo-labels. These enhancements enabled our approach to secure the 3rd place in the challenge, achieving a 3D AP of 62.67% for the Car category on the phase-1 target domain, and 58.76% and 49.81% for Car and Pedestrian categories respectively on the phase-2 target domain.
Paper Structure (10 sections, 1 figure, 1 table)

This paper contains 10 sections, 1 figure, 1 table.

Table of Contents

  1. Introduction
  2. Method
  3. Cross-domain 3D object detection paradigm
  4. Data Augmentation
  5. Training framework
  6. Experiments
  7. Experimental Setup
  8. Implementation Details
  9. Ablation Study
  10. Conclusion

Figures (1)

  • Figure 1: Overall pipeline of the proposed Cross-Platform 3D Detector. Source domain point clouds are processed by PVRCNN++ to produce detection results with labels, while target domain data generates pseudo-labels through the same detector. The ST3D module refines these pseudo-labels, and the CJA module operates exclusively on the source domain to enhance robustness by simulating platform-specific viewpoint jitter. Both source labels and refined target pseudo-labels are then utilized to iteratively update the model, enabling progressive adaptation across platforms without additional annotations.