PaperlayerView exact revision on arXiv

Source-linked AI summary

PC-RGNN: Point Cloud Completion and Graph Neural Network for 3D Object Detection

Yanan Zhang, Di Huang, Yunhong Wang

arXiv:2012.10412v3cs.CV

TL;DR

Distant and occluded objects produce sparse or partial LiDAR point clouds that challenge 3D detection. PC-RGNN combines point cloud completion with an attention-based multi-scale GNN, and the authors report better KITTI results than other approaches, especially for Moderate and Hard cases.

  • Problem

    Distant and occluded objects yield sparse and partial point clouds, while existing detectors struggle with sampling variation and insufficient use of the input.

    Source: §1 · p. 1

  • Method

    PC-RGNN generates 3D proposals, completes sparse or partial proposals, and uses AMS-GNN with local-global attention and multi-scale graph aggregation to encode geometric relations.

    Source: §1 · pp. 1–2

  • Results

    On the KITTI validation set, PC-RGNN outperforms the other approaches, especially in Moderate and Hard difficulties.

    Source: §4.2 · p. 6

  • Takeaways & Limitations

    The authors conclude that PC-RGNN is effective for KITTI detection with sparse and partial point-cloud inputs.

    Source: §5 · p. 7

Abstract

from arXiv · show

LiDAR-based 3D object detection is an important task for autonomous driving and current approaches suffer from sparse and partial point clouds of distant and occluded objects. In this paper, we propose a novel two-stage approach, namely PC-RGNN, dealing with such challenges by two specific solutions. On the one hand, we introduce a point cloud completion module to recover high-quality proposals of dense points and entire views with original structures preserved. On the other hand, a graph neural network module is designed, which comprehensively captures relations among points through a local-global attention mechanism as well as multi-scale graph based context aggregation, substantially strengthening encoded features. Extensive experiments on the KITTI benchmark show that the proposed approach outperforms the previous state-of-the-art baselines by remarkable margins, highlighting its effectiveness.

1 Introduction

LiDAR point clouds provide 3D geometry, but distant and occluded objects yield sparse or incomplete inputs that challenge existing detectors. PC-RGNN addresses these inputs with point cloud completion and an attention-based multi-scale GNN.

1 Introduction

  • Existing detectors can miss distant and occluded objects because point-cloud representations handle sampling variation poorly and PointNet-like backbones underuse the input.LiDAR point clouds convey 3D geometry and spatial locations and are less sensitive to illumination variations than 2D images.Source: §1 · p. 1
  • PC-RGNN first generates 3D proposals, then completes sparse or partial proposals and encodes their shape with AMS-GNN.AMS-GNN combines multi-scale graphs with local-global attention to capture geometric relations and strengthen features.Source: §1 · pp. 1–2
  • The authors report that PC-RGNN significantly boosts detection performance and reaches competitive KITTI scores, including in difficult scenes.The introduction credits completion-based data refinement and attention-based multi-scale GNN feature strengthening.Source: §1 · p. 2
  • The authors present point cloud completion as the first use of point-cloud refinement in 3D object detection.The module is intended to improve proposals containing sparse and partial points.Source: §1 · p. 2

2 Related Work

Prior 3D detectors extract features from grids, raw points, or graphs, while point-cloud completion methods reconstruct shapes from partial inputs. PC-RGNN distinguishes its approach by applying completion to LiDAR-based 3D detection.

2 Related Work

  • Grid-based detectors project or voxelize point clouds for CNN processing, but coarse quantization loses information and limits performance.Examples include view-based methods (Chen et al. 2017; Ku et al. 2018) and voxel-based methods (Zhou and Tuzel 2018; Yan, Mao, and Li 2018).Source: §2 · p. 2
  • Point-based detectors outperform grid-based ones, yet lack of explicit point-relation modeling limits discriminative geometry representation.Examples include PointNet-based methods (Qi et al. 2018; Shi, Wang, and Li 2019).Source: §2 · p. 2
  • Point-GNN uses graph neural networks for detection, but equal node contributions and a single-scale graph leave contextual information underused (Shi and Rajkumar 2020).The passage identifies both design choices as opportunities for improvement.Source: §2 · p. 2
  • Prior completion methods estimate full shapes from partial clouds; PC-RGNN applies completion to LiDAR detection of distant or seriously occluded objects.The authors describe this as the first attempt to use completion to ameliorate LiDAR-based 3D object detection (Yuan et al. 2018).Source: §2 · pp. 2–3

3 PC-RGNN

PC-RGNN is a three-module pipeline: it generates 3D proposals, completes sparse or partial point clouds, and applies an attention-based multi-scale GNN to predict detections. Its completion module adds points while preserving the input arrangement, and its GNN aggregates local and global context.

3.1 Framework

  • The framework segments foreground from background and generates bounding-box candidates, then completes proposal shapes and encodes their structure to predict detections.The completion module recovers dense, entire 3D shapes from sparse or partial proposals.Source: §3.1 · p. 3

3.2 3D Proposal Generation

  • The proposal-generation module learns point-wise features to locate foreground areas and generate 3D bounding-box candidates in a bottom-up manner.It avoids a large set of predefined 3D anchors, reducing the search space in this phase.Source: §3.2 · p. 3

3.3 Point Cloud Completion

  • The PC module supplements sparse or partial proposals with additional points while leaving the input data unchanged to preserve its spatial arrangement.Its architecture uses a Multi-Resolution Graph Encoder, Point Pyramid Decoder, and Discriminator Network.Source: §3.3 · p. 3, Figure 3 · p. 4

3.4 Attention based Multi-Scale GNN

  • AMS-GNN combines multi-scale graph aggregation with local attention and global attention to strengthen proposal features for detection.The module contains four AMS-GCN layers and a global attention layer.Source: §3.4 · pp. 4–5

4 Experiments

The experiments evaluate PC-RGNN on KITTI, report performance on the test and validation splits, and use ablations to assess its point-cloud completion and AMS-GNN components.

4.1 Datasets and Implementation Details

  • KITTI evaluation uses 3,712 training samples and 3,769 validation samples, with ShapeNetCars pretraining and KITTI-object fine-tuning for point-cloud completion.The KITTI benchmark has 7,481 training point clouds and 7,518 test point clouds; the completion module is fine-tuned using 2,000 KITTI object clouds with more than 2,048 points.Source: §4.1 · p. 5

4.2 Results

  • PC-RGNN significantly outperforms previously published state-of-the-art counterparts across all KITTI test tasks and difficulty levels.The authors identify Point-GNN and PointRCNN as baselines because they represent graph-based detection and a two-stage approach, respectively.Source: §4.2 · p. 6
  • On the KITTI validation split, PC-RGNN outperforms other approaches by remarkable margins, especially at Moderate and Hard difficulty.Table 2 reports car-class performance; the car IoU threshold is 0.7.Source: §4.2 · p. 6, Table 2 · p. 6

4.3 Ablation Study

  • Adding PC alone or AMS-GNN alone raises 3D mAP from the 81.63% baseline to 82.54% and 83.18%, respectively; together they reach 84.27%, a 2.64% gain.Table 3 presents the ablation of the two modules.Source: §4.3 · p. 6, Table 3 · p. 6
  • The AMS-GNN ablation reports improvements of 0.21%, 0.62%, and 0.16% from multi-scale aggregation, local attention, and global attention, respectively.The authors report that all three components improve the GNN's ability to capture point-cloud geometric characteristics.Source: §4.3 · p. 6, Table 4 · p. 6
  • Removing graph encoding or the multi-resolution branch from PC-M lowers performance by 0.60% or 0.67%, while replacing PC-M with PC-O lowers it by 0.86%.The authors conclude that original points remain important for regressing object locations and that PC-M handles this issue better than PC-O.Source: §4.3 · pp. 6–7

5 Conclusion

PC-RGNN combines point-cloud completion and graph-based feature enhancement for LiDAR 3D detection. The paper reports state-of-the-art results on KITTI.

5 Conclusion

  • The paper includes ablation experiments on the point-cloud completion module.Table 5 is identified as the PC-module ablation table, but the supplied passage gives no outcomes.Source: Table 5 · p. 7
  • PC-RGNN reaches state-of-the-art results on the KITTI dataset.The passage reports no specific metrics or values.Source: §5 · p. 7
  • The point-cloud completion module refines sparse, partial proposals; the paper describes this as the first attempt to integrate completion into 3D object detection.Source: §5 · p. 7
  • The graph neural network enhances features by capturing geometric relations through local-global attention and multi-scale graph-based context aggregation.Source: §5 · p. 7

Related papers

Loading 2012.10412v3…