PointAugment.pdf
需积分: 10 85 浏览量
2020-03-21
18:28:01
上传
评论
收藏 1.86MB PDF 举报
PointAugment: an Auto-Augmentation Framework
for Point Cloud Classification
Ruihui Li
1
Xianzhi Li
1,2
Pheng-Ann Heng
1,2
Chi-Wing Fu
1,2
1
The Chinese University of Hong Kong
2
Guangdong Provincial Key Laboratory of Computer Vision and Virtual Reality Technology,
Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, China
{lirh,xzli,pheng,cwfu}@cse.cuhk.edu.hk
Abstract
We present PointAugment
1
, a new auto-augmentation
framework that automatically optimizes and augments point
cloud samples to enrich the data diversity when we train
a classification network. Different from existing auto-
augmentation methods for 2D images, PointAugment is
sample-aware and takes an adversarial learning strategy to
jointly optimize an augmentor network and a classifier net-
work, such that the augmentor can learn to produce aug-
mented samples that best fit the classifier. Moreover, we
formulate a learnable point augmentation function with a
shape-wise transformation and a point-wise displacement,
and carefully design loss functions to adopt the augmented
samples based on the learning progress of the classifier. Ex-
tensive experiments also confirm PointAugment’s effective-
ness and robustness to improve the performance of various
networks on shape classification and retrival.
1. Introduction
In recent years, there has been a growing interest in de-
veloping deep neural networks [20, 21, 33, 16, 15] for 3D
point clouds. To robustly train a network often relies on the
availability and diversity of the data. However, unlike 2D
image benchmarks such as ImageNet [8] and MS COCO
dataset [12], which have over millions of training samples,
3D datasets are typically much smaller in quantity, with rel-
atively small amount of labels and limited diversity. For
instance, ModelNet40 [34], one of the most commonly-
used benchmark for 3D point cloud classification, only has
12,311 models of 40 categories. The limited data quantity
and diversity may cause overfitting problem and further af-
fect the generalization ability of the network.
Nowadays, data augmentation (DA) is a very common
strategy to avoid overfitting and improve the network gener-
1
https://github.com/liruihui/PointAugment
Figure 1: Classification accuracy (%) on ModelNet40 with
or without training the networks with our PointAugment.
We can see clear improvements on four representative net-
works. More comparison results are presented in Section 5.
alization ability by artificially enlarging the quantity and di-
versity of the training samples. For 3D point clouds, due to
the limited amount of training samples and an immense aug-
mentation space in 3D, conventional DA strategies [20, 21]
often simply perturb the input point cloud randomly in a
small and fixed pre-defined augmentation range to main-
tain the class label. Despite its effectiveness for the existing
classification networks, this conventional DA approach may
lead to insufficient training, as summarized below.
First, existing methods regard the network training and
DA as two independent phases without jointly optimizing
them, e.g., feedback the training results to enhance the DA.
Hence, the trained network could be suboptimal. Second,
existing methods apply the same fixed augmentation pro-
cess with rotation, scaling, and/or jittering, to all input point
cloud samples. The shape complexity of the samples is ig-
nored in the augmentation, e.g., a sphere remains the same
no matter how we rotate it, but a complex shape may need
larger rotations. Hence, conventional DA may be redundant
or insufficient for augmenting the training samples [5].
1
arXiv:2002.10876v1 [cs.CV] 25 Feb 2020
资源评论
