17 天前

PointNet++:度量空间中点集的深度分层特征学习

PointNet++:度量空间中点集的深度分层特征学习

摘要

此前很少有研究探讨点集上的深度学习问题。Qi等人提出的PointNet是该领域的开创性工作。然而,由于其设计限制,PointNet无法捕捉由度量空间所诱导的局部结构,从而限制了其识别细粒度模式的能力,以及在复杂场景中的泛化性能。在本工作中,我们提出一种分层神经网络,通过在输入点集的嵌套划分上递归应用PointNet,实现对点集的多层次建模。通过利用度量空间中的距离信息,我们的网络能够逐步学习具有越来越广泛上下文感知能力的局部特征。此外,我们观察到,点集通常以不同密度进行采样,而基于均匀密度训练的网络在面对非均匀采样时性能显著下降。为此,我们设计了新型的集合学习层,可自适应地融合多尺度特征。实验表明,我们提出的网络——PointNet++,能够高效且稳健地学习深层点集特征。尤其在具有挑战性的三维点云基准测试中,其性能显著优于现有最先进方法。

代码仓库

AsahiLiu/PointDetectron
pytorch
GitHub 中提及
ikh-innovation/roboweldar-votenet
pytorch
GitHub 中提及
caizeyu1992/pointnet2
mindspore
GitHub 中提及
Lw510107/PointNet
tf
GitHub 中提及
hehefan/PointRNN-PyTorch
pytorch
GitHub 中提及
ftdlyc/pointnet_pytorch
pytorch
GitHub 中提及
referit3d/referit3d
pytorch
GitHub 中提及
johnsonsign/mast-pre
pytorch
GitHub 中提及
hehefan/P4Transformer
pytorch
GitHub 中提及
LONG-9621/PointNet-
tf
GitHub 中提及
witignite/Frustum-PointNet
tf
GitHub 中提及
zhh6425/LocalContextPropagation
pytorch
GitHub 中提及
sshaoshuai/Pointnet2.PyTorch
pytorch
GitHub 中提及
facebookresearch/votenet
pytorch
GitHub 中提及
llzlcl/pointcloud-segment
tf
GitHub 中提及
nickgkan/beauty_detr
pytorch
GitHub 中提及
iballester/SPiKE
pytorch
GitHub 中提及
zaiweizhang/H3DNet
pytorch
GitHub 中提及
timothylimyl/PointNet-Pytorch
pytorch
GitHub 中提及
rusty1s/pytorch_cluster
pytorch
GitHub 中提及
open-air-sun/pq-transformer
pytorch
GitHub 中提及
curryyuan/x-trans2cap
pytorch
GitHub 中提及
brbzjl/pointnet2
tf
GitHub 中提及
facebookresearch/imvotenet
pytorch
GitHub 中提及
Harut0726/votenet
pytorch
GitHub 中提及
xurui1217/pointnet2-master
tf
GitHub 中提及
thu17cyz/3DIoUMatch
pytorch
GitHub 中提及
joyhsu0504/ns3d
pytorch
GitHub 中提及
asalarpour/Point_GN
pytorch
GitHub 中提及
hehefan/PSTNet2
pytorch
GitHub 中提及
Lonepic/SPIB
pytorch
GitHub 中提及
JohnsonSign/PointCMP
pytorch
GitHub 中提及
tingxueronghua/dpke
pytorch
GitHub 中提及
tonysy/pointnet2_tf
tf
GitHub 中提及
LONG-9621/VoteNet
pytorch
GitHub 中提及
houseleo/pointnet
tf
GitHub 中提及
charlesq34/pointnet2
tf
GitHub 中提及
hoangcuongbk80/VoteGrasp
pytorch
GitHub 中提及
nickgkan/butd_detr
pytorch
GitHub 中提及
zyang-ur/SAT
pytorch
GitHub 中提及
hehefan/PointRNN
tf
GitHub 中提及
zhh6425/MotionPointNet
pytorch
GitHub 中提及
zenroad/modifypointnet
tf
GitHub 中提及
dfki-av/mikasa-3dvg
pytorch
GitHub 中提及
FlowWind1999/pointnet-2
tf
GitHub 中提及
Xiangxu-0103/Octant-CNN
tf
GitHub 中提及

基准测试

基准方法指标
3d-part-segmentation-on-intraPointNet++
DSC (A): 84.64
DSC (V): 96.48
IoU (A): 76.38
IoU (V): 93.42
3d-part-segmentation-on-shapenet-partPointNet++
Class Average IoU: 81.9
Instance Average IoU: 85.1
3d-point-cloud-classification-on-intraPointNet++
F1 score (5-fold): 0.903
3d-point-cloud-classification-on-modelnet40PointNet++
Number of params: 1.74M
Overall Accuracy: 90.7
3d-point-cloud-classification-on-modelnet40-cPointNet++
Error Rate: 0.236
3d-point-cloud-classification-on-scanobjectnnPointNet++
Mean Accuracy: 75.4
OBJ-BG (OA): 82.3
OBJ-ONLY (OA): 84.3
Overall Accuracy: 77.9
3d-semantic-segmentation-on-dalesPointNet++
Model size: 3.0M
Overall Accuracy: 95.7
mIoU: 68.3
3d-semantic-segmentation-on-kitti-360PointNet++
Model size: 3.0M
mIoU Category: 58.28
miou: 35.66
3d-semantic-segmentation-on-scannet-1PointNet++
Top-1 IoU: 0.201
Top-3 IoU: 0.389
3d-semantic-segmentation-on-semantickittiPointNet++
test mIoU: 20.1%
3d-semantic-segmentation-on-stpls3dPointNet++
mIOU: 15.92
few-shot-3d-point-cloud-classification-on-1PointNet++
Overall Accuracy: 38.53
Standard Deviation: 16.0
few-shot-3d-point-cloud-classification-on-2PointNet++
Overall Accuracy: 42.39
Standard Deviation: 14.2
few-shot-3d-point-cloud-classification-on-3PointNet++
Overall Accuracy: 23.05
Standard Deviation: 7.0
few-shot-3d-point-cloud-classification-on-4PointNet++
Overall Accuracy: 18.80
Standard Deviation: 7.0
person-re-identification-on-dukemtmc-reidPointNet++ (MSG) [qi2017pointnet++]
Rank-1: 60.23
mAP: 39.36
point-cloud-segmentation-on-pointcloud-cPointNet++
mean Corruption Error (mCE): 1.112
semantic-segmentation-on-scannetPointNet++
test mIoU: 33.9
val mIoU: 53.5
semantic-segmentation-on-shapenetPointNet++
Mean IoU: 84.6%
semantic-segmentation-on-toronto-3d-l002PointNet++
mIoU: 56.5
oAcc: 91.2
supervised-only-3d-point-cloud-classificationPointNet++
GFLOPs: 1.7
Number of params (M): 1.5
Overall Accuracy (PB_T50_RS): 77.9

用 AI 构建 AI

从想法到上线——通过免费 AI 协同编程、开箱即用的环境和市场最优价格的 GPU 加速您的 AI 开发

AI 协同编程
即用型 GPU
最优价格
立即开始

Hyper Newsletters

订阅我们的最新资讯
我们会在北京时间 每周一的上午九点 向您的邮箱投递本周内的最新更新
邮件发送服务由 MailChimp 提供
PointNet++:度量空间中点集的深度分层特征学习 | 论文 | HyperAI超神经