PTNET - A Proposal-centric Transformer Network for 3D Object Detection - 论文摘要与方法概述

Sep 10, 2026

Uploaded by William

3D object detection with LiDAR point clouds is vital for autonomous driving perception, yet existing two-stage models often underperform because proposal quality is weak. The generated proposal features suffer geometric detail degradation from sparsity and uneven point distributions, and they fail to use contextual cues during independent refinement, losing complementary information from adjacent proposals. PTN introduces a proposal-centric transformer with HAFA for hierarchical attentive feature alignment and CPRM for collaborative proposal refinement. Experiments on Waymo and KITTI confirm PTN’s superiority.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages17

Reading Time43 min

AccessFree to download

Abstract

Method Overview

Introduction

Proposal Quality Limitations

Geometric Detail Degradation

Contextual Cue Inefficiency

PTN Design: HAFA and CPRM

PTNET: A PROPOSAL-CENTRIC TRANSFORMER NETWORK FOR 3D OBJECT DETECTION Jianping Zhong 1 Zhaobo Qi 1∗ Kaiwen Duan2 Xinyan Liu 1 Beichen Zhang 1 Weigang Zhang 1 ,3 ∗ Qingming Huang2 1Harbin Institute of Technology, Weihai 2University of Chinese Academy of Sciences 3Harbin Institute of Technology (Weihai) Qingdao Research Institute [jianpingzhong@stu.hit.edu.cn](jianpingzhong@stu.hit.edu.cn), [kaiwenduan@outlook.com](kaiwenduan@outlook.com),{qizb, xinyliu, beiczhang, [wgzhang](wgzhang}@hit.edu.cn)[}](wgzhang}@hit.edu.cn)[@hit.edu.cn](wgzhang}@hit.edu.cn), [qmhuang@ucas.ac.cn](qmhuang@ucas.ac.cn) ABSTRACT 3D object detection using LiDAR point cloud data is critical for autonomous driving systems. However, recent two-stage detectors still struggle to deliver satisfactory performance primarily due to inadequate proposal quality, which stems from significant geometric detail degradation in generated proposal features caused by high sparsity and uneven distribution of point clouds, as well as a complete failure to exploit surrounding contextual cues during independent proposal refinement, losing complementary details from adjacent proposals. To this end, we propose a Proposal-centric Transformer Network (PTN), which includes a Hierarchical Attentive Feature Alignment (HAFA) and a Collaborative Proposal Refinement Module (CPRM) . More concretely, HAFA employs a dual-stream architecture to extract multi-granularity proposal representations, including coarse-grained multi-scale voxel features and fine-grained coordinate point features to enhance proposals’ object geometric representation ability. CPRM first generates hybrid object queries for all objects and then establishes contextual-aware interactions through the 3D parameter-guided deformable attention mechanism to effectively aggregate spatial location cues and category-specific information across proposals that are spatially adjacent and semantically correlated. Extensive experiments on the large-scale Waymo and KITTI benchmarks demonstrate the superiority of PTN. The code is available at [https://github.com/ZhongJianPing1/](https://github.com/ZhongJianPing1/)[ ](https://github.com/ZhongJianPing1/)[ptnet.git](ptnet.git.)[.](ptnet.git.) 1 INTRODUCTION 3D object detection serves as a foundational task for environmental perception in autonomous driving, aiming at precise localization and classification of objects in 3D scenes. Recently, to achieve a trade-off between performance and efficiency, researchers have focused on the two-stage 3D object detector paradigm Deng et al. (2021); Shi et al. (2023): this paradigm first employs the region proposal network (RPN) to obtain proposals, then extracts proposal features via ROI pooling Denget al. (2021), and ultimately produces detection outputs by refining these proposal features in the subsequent refinement stage. However, existing two-stage 3D object detectors are constrained by suboptimal proposal quality, which arises from two key underlying issues. The primary concern revolves around the degradation of geometric detail in the generated proposal features. In particular, prevailing two-stage detectors typically progressively expand receptive fields through pooling operations to generate proposal features. For areas of objects with few points, the pooling operation tends to filter out high-frequency geometric features such as surface details and edge sharpness, resulting in blurred proposal boundaries and structural incompleteness. As shown in Figure 1, certain objects have only a few points or sparsely distributed points. After multi-pooling operations, the detailed information inherent in these points is lost, leading to inaccurate predictions (denoted by green boxes) . Although recent research has explored the use of foreground-agnostic ∗ Corresponding Authors. Figure 1: Qualitative results under the bird’s-eye view (BEV) on Waymo. The red and black 3D boxes are annotated by humans, and the red points represent th

FAQs

Why do current two-stage 3D object detectors struggle in proposal quality?

They degrade geometric details in proposal features due to high sparsity and uneven point distributions, and they optimize proposals independently, failing to exploit surrounding contextual cues from adjacent proposals.

What does PTN propose to improve proposal representation?

PTN uses HAFA, a hierarchical attentive feature alignment module with a dual-stream architecture to capture multi-granularity proposal representations from coarse multi-scale voxel features and fine coordinate point features.

How does PTN refine proposals using context?

PTN’s CPRM first generates hybrid object queries, then builds contextual-aware interactions using a 3D parameter-guided deformable attention mechanism to aggregate spatial-location cues and category-specific information across spatially adjacent, semantically correlated proposals.

Sep 10, 2026

Uploaded by William