PseudoAugment Learning to Use Unlabeled Data for Data Augmentation in Point Clouds Zhaoqi Leng1 Shuyang Cheng1 Benjamin Caine2 Weiyue Wang1 Xiao

2025-05-02 0 0 4.92MB 19 页 10玖币

侵权投诉

PseudoAugment: Learning to Use Unlabeled

Data for Data Augmentation in Point Clouds

Zhaoqi Leng1, Shuyang Cheng1, Benjamin Caine2, Weiyue Wang1, Xiao

Zhang1, Jonathon Shlens2, Mingxing Tan1, and Dragomir Anguelov1

Waymo1, Google2

1{lengzhaoqi, weiyuewang, tanmingxing, dragomir}@waymo.com

Abstract. Data augmentation is an important technique to improve

data eﬃciency and save labeling cost for 3D detection in point clouds.

Yet, existing augmentation policies have so far been designed to only

utilize labeled data, which limits the data diversity. In this paper, we

recognize that pseudo labeling and data augmentation are complemen-

tary, thus propose to leverage unlabeled data for data augmentation to

enrich the training data. In particular, we design three novel pseudo-

label based data augmentation policies (PseudoAugments) to fuse both

labeled and pseudo-labeled scenes, including frames (PseudoFrame), ob-

jects (PseudoBBox), and background (PseudoBackground). PseudoAug-

ments outperforms pseudo labeling by mitigating pseudo labeling er-

rors and generating diverse fused training scenes. We demonstrate Pseu-

doAugments generalize across point-based and voxel-based architectures,

diﬀerent model capacity and both KITTI and Waymo Open Dataset. To

alleviate the cost of hyperparameter tuning and iterative pseudo label-

ing, we develop a population-based data augmentation framework for 3D

detection, named AutoPseudoAugment. Unlike previous works that per-

form pseudo-labeling oﬄine, our framework performs PseudoAugments

and hyperparameter tuning in one shot to reduce computational cost.

Experimental results on the large-scale Waymo Open Dataset show our

method outperforms state-of-the-art auto data augmentation method

(PPBA) and self-training method (pseudo labeling). In particular, Au-

toPseudoAugment is about 3×and 2×data eﬃcient on vehicle and

pedestrian tasks compared to prior arts. Notably, AutoPseudoAugment

nearly matches the full dataset training results, with just 10% of the

labeled run segments on the vehicle detection task.

Keywords: data augmentation, semi-supervised learning, 3D detection.

1 Introduction

3D object detection from LiDAR point cloud data is a core component of au-

tonomous driving. Building accurate 3D object detection systems requires vast

quantities of labeled scenes with accurate 3D bounding box annotations. While

unlabeled LiDAR data is readily available, labeling itself is costly, e.g., 6.4 hours

of LiDAR data contains more than 10 million human labeled 3D boxes [58].

arXiv:2210.13428v1 [cs.CV] 24 Oct 2022

2 Z. Leng et al.

Fig. 1: AutoPseudoAugment is more data eﬃcient than auto data aug-

mentation and self-training methods. Data augmentation only (PPBA [8]),

self-training (Pseudo labeling [4]) and our method (AutoPseudoAugment) are

evaluated using 3D detection AP at Level 1 diﬃculty on the validation split of

the Waymo Open Dataset [58]. Using 10% of labeled run segments, AutoPseu-

doAugment is about 3×data eﬃcient as PPBA and Pseudo label method on the

vehicle class and 2×on the pedestrian class. AutoPseudoAugment is nearly 10×

and more than 5×data eﬃcient compared to the supervised (no augmentation)

vehicle and pedestrian baselines.

Thus, an eﬀective way to increase the data eﬃciency for model training would

be very appealing.

Data augmentation is an eﬀective way to increase data eﬃciency for labeled

data. Data augmentations for 3D detection generally come in two forms: global

augmentations like scene rotations, or local augmentations like ground truth aug-

mentation, where crops of ground truth objects from the training set are inserted

into the scene. Pasting ground truth objects into the scene has been shown to

be extremely eﬀective on various 3D detection datasets [64,25,8,66,70,65,30].

However, these augmentation techniques are typically limited to the labeled

training data. A simple way to incorporate unlabeled data into training is pseudo

labeling, but naively applying existing 3D data augmentation policies to pseudo

labeled frames has an intrinsic limitation, i.e., pseudo labeled frames contain nu-

merous false positive/negative bounding boxes and points. Several recent studies

on 3D pseudo labeling [4,44] have tried to use large-capacity teacher models to

mitigate this issue, but the intrinsic pseudo-labeling errors persist. Here, we seek

to an alternative approach: mitigating the pseudo labeled errors by new data aug-

mentation policies.

Another challenge is how to eﬀectively combine labeled and unlabeled data

via data augmentation. Previous approaches treat pseudo-labeled frames as

a whole and do not recognize the compositional nature of 3D point clouds

scenes [4,44]. This limits the diversity of training data. A simple way to fuse la-

PseudoAugment 3

PseudoFrame PseudoBBox PseudoBackground

Labeled frame Pseudo labeled frame

Augmented frames

Original PseudoAugments

Fig. 2: Visualization of PseudoAugments. PseudoAugments contain three

new data augmentation policies: PseudoFrame, PseudoBBox, and PseudoBack-

ground. PseudoFrame replaces the labeled frame with a pseudo-labeled frame

and drops points of low-conﬁdence bounding boxes in the pseudo-labeled frames.

PseudoBBox pastes high-conﬁdence bounding boxes and corresponding point

clouds from a pseudo-labeled frame to a labeled frame. PseudoBackground re-

moves all points within bounding boxes in a pseudo-labeled frame, and replaces

the background point clouds in the labeled frame with the background point

clouds of the pseudo-labeled frame. The augmented frames are used as labeled

frames during training.

beled and pseudo-labeled frame is to generalize the existing copy-pasting object

data augmentation to leverage unlabeled objects. Interestingly, we observe that

only pasting objects between labeled and pseudo-labeled frames is not enough

[64,25,8,66,70,65,30], because we miss out the diverse background scenes in the

pseudo labeled dataset. Especially for 3D point clouds, more than 90% of the

points are backgrounds, which provide critical ingredients for 3D detectors to

learn to detect objects in new scenes. Thus, it is necessary to develop a set of data

augmentation policies that take advantage of both foreground objects and back-

ground points in the pseudo labeled frames along with labeled frames to generate

combinatorial number of point clouds.

In this work, we propose a set of data augmentation policies tailored for

pseudo labeled data, named PseudoAugments. As shown in Figure 2, our Pseu-

doAugments contain three new data augmentation policies: PseudoFrame re-

moves low conﬁdence points, PseudoBBox pastes pseudo objects onto labeled

scenes, and PseudoBackground swaps the background point clouds between la-

4 Z. Leng et al.

beled and pseudo-labeled scenes. All our augmentations allow pseudo-labeling

uncertainty, and only make use of points of frame, object, and background with

high-conﬁdence. PseudoAugments signiﬁcantly increase the diversity of training

data by enabling a combinatorial number of new fused training scenes, including

1) ground truth objects on pseudo labeled background scenes, 2) pseudo labeled

objects on ground truth background scenes, and 3) pseudo labeled objects on

pseudo labeled background scenes, which greatly enrich the diversity of training

data.

Based on PseudoAugments, we develop an auto data augmentation frame-

work named AutoPseudoAugment to learn the best augmentation policies. Our

AutoPseudoAugment is based on population-based training (PBT) and searches

for the best augmentation policies online at diﬀerent training stages. On top of

PBT, AutoPseudoAugment uses the top-performing models in previous genera-

tions as an ensemble of teachers to pseudo label unlabeled data, which further

boosts the quality of pseudo labeled data without the need of training a sepa-

rated set of high-capacity teacher models [4,44]. AutoPseudoAugment extends

PBT beyond simple hyperparameter tuning by introducing population-based

distillation and creates a virtuous cycle between students and teachers, where

good teachers in previous generations improves the quality of student models,

which become better teachers to pseudo label for future generations.

Our main contributions can be summarized as follows:

1. PseudoAugments: unifying data augmentation and pseudo label-

ing. We identify data augmentation and pseudo labeling are complementary

and introduce PseudoFrame, PseudoBBox, PseudoBackground data aug-

mentation policies to take advantage of the composability of unlabeled 3D

point clouds while mitigating errors.

2. AutoPseudoAugment: eﬃcient one-shot framework for PseudoAug-

ment. Our framework extends PBT by introducing population-based dis-

tillation. AutoPseudoAugment does auto hyperparameters search and self-

training in one-shot, which reduces the training cost.

3. Extensive experimental evaluations. We demonstrate PseudoAugments

generalize to diﬀerent network architectures, model sizes, and datasets. In

addition, AutoPseudoAugment outperforms both state-of-the-art auto data

augmentation method (PPBA [8]) and pseudo labeling [4]. In particular,

leveraging unlabeled data, AutoPseudoAugment requires 10% of labeled run

segments to achieve similar performance as PPBA training on 30% of run

segments and nearly matches the model performance trained on all labeled

data without data augmentation, shown in Figure 1.

2 Related Work

2.1 Data augmentation

Data augmentation has been widely adopted to improve the performance of mod-

els trained with supervised learning, such as image classiﬁcation[55,10,60,49,29,13,67],

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

PseudoAugment:LearningtoUseUnlabeledDataforDataAugmentationinPointCloudsZhaoqiLeng1,ShuyangCheng1,BenjaminCaine2,WeiyueWang1,XiaoZhang1,JonathonShlens2,MingxingTan1,andDragomirAnguelov1Waymo1,Google21flengzhaoqi,weiyuewang,tanmingxing,dragomirg@waymo.comAbstract.Dataaugmentationisanimportanttechniqu...

展开>> 收起<<

PseudoAugment Learning to Use Unlabeled Data for Data Augmentation in Point Clouds Zhaoqi Leng1 Shuyang Cheng1 Benjamin Caine2 Weiyue Wang1 Xiao.pdf

共19页,预览4页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

PseudoAugment Learning to Use Unlabeled Data for Data Augmentation in Point Clouds Zhaoqi Leng1 Shuyang Cheng1 Benjamin Caine2 Weiyue Wang1 Xiao

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: