Voxel Densification for Serialized 3D Object Detection: Mitigating Sparsity via Pre-serialization Expansion

This repository serves as the official implementation for the paper:
Voxel Densification for Serialized 3D Object Detection: Mitigating Sparsity via Pre-serialization Expansion

Abstract: Recent advances in point cloud object detection have increasingly adopted Transformer-based and State Space Models (SSMs) to capture long-range dependencies. However, these serialized frameworks strictly maintain the consistency of input and output voxel dimensions, inherently lacking the capability for voxel expansion. This limitation hinders performance, as expanding the voxel set is known to significantly enhance detection accuracy, particularly for sparse foreground objects. To bridge this gap, we propose a novel Voxel Densification Module (VDM). Unlike standard convolutional stems, VDM is explicitly designed to promote pre-serialization spatial expansion. It leverages sparse 3D convolutions to propagate foreground semantics to neighboring empty voxels, effectively densifying the feature representation before it is flattened into a sequence. VDM serves two key functions: (1) enhancing spatial connectivity via voxel densification, and (2) aggregating fine-grained local context through residual sparse blocks. Crucially, to balance the computational overhead of increased voxel density, we introduce a strategic downsampling mechanism. We integrate VDM into both Transformer-based (DSVT) and SSM-based (LION) detectors. Extensive experiments demonstrate that VDM consistently improves detection accuracy across multiple benchmarks.

🚀 News

• [2026-02-05] Initial code release.

✨ Highlights

• Pre-serialization Spatial Expansion: Defines a new paradigm to explicitly expand the foreground voxel set before sequence flattening, addressing the sparsity limitation in serialized models.
• Generic Plugin: Seamlessly integrates with state-of-the-art serialized detectors, including Transformer-based (DSVT) and SSM-based (LION) frameworks.

🏆 Model Zoo & Main Results

We provide the checkpoints and logs for our main models, including the full VDM and the Only-Densification (VDM-OD) variant.

Model	Dataset	Split	Metric	Performance	Config	Baidu Pan	Hugging Face
VDM	Waymo	Val	L2 mAPH	74.8	waymo_vdm.yaml	Baidu Pan	HF
VDM-OD	Waymo	Val	L2 mAPH	74.8	waymo_vdm_od.yaml	Baidu Pan	HF
VDM	nuScenes	Val	mAP	68.1	nuscenes_vdm.yaml	Baidu Pan	HF
VDM-OD	nuScenes	Val	mAP	68.5	nuscenes_vdm_od.yaml	Baidu Pan	HF
VDM	Argoverse 2	Val	mAP	42.3	argo2_vdm.yaml	Baidu Pan	HF
VDM-OD	Argoverse 2	Val	mAP	42.6	argo2_vdm_od.yaml	Baidu Pan	HF
VDM	ONCE	Val	mAP	67.6	once_vdm.yaml	Baidu Pan	HF
VDM-OD	ONCE	Val	mAP	66.1	once_vdm_od.yaml	Baidu Pan	HF

Note for Baidu Pan links: > - VDM extraction code: jk22

• VDM-OD extraction code: nb3f

Note for Hugging Face links: > - You can also browse the full model repository directly at hfffkk/VDM.

🛠️ Getting Started

Environment Requirements

This code has been tested in the following environment. Other versions might also work, but we recommend matching these for the best compatibility:

• OS: Linux (Ubuntu)
• Python: 3.10
• CUDA: 11.8
• PyTorch: 2.1.0
• Spconv: 2.3.6 (spconv-cu118)

Installation

Please refer to docs/INSTALL.md for detailed step-by-step installation instructions, including environment setup and dependency installation.

Quick Start

Please refer to docs/GETTING_STARTED.md to learn how to prepare the datasets and run the training and testing scripts.

🙏 Acknowledgements

Our codebase integrates and builds upon the LION framework and OpenPCDet. We sincerely thank the authors for their outstanding work:

@article{liu2024lion,
  title={LION: Linear Group RNN for 3D Object Detection in Point Clouds},
  author={Zhe Liu, Jinghua Hou, Xingyu Wang, Xiaoqing Ye, Jingdong Wang, Hengshuang Zhao, Xiang Bai},
  journal={Advances in Neural Information Processing Systems},
  year={2024}
}