没有合适的资源?快使用搜索试试~ 我知道了~
Hierarchical Aggregation for 3D Instance Segmentation.pdf
1.该资源内容由用户上传,如若侵权请联系客服进行举报
2.虚拟产品一经售出概不退款(资源遇到问题,请及时私信上传者)
2.虚拟产品一经售出概不退款(资源遇到问题,请及时私信上传者)
版权申诉
0 下载量 178 浏览量
2021-11-21
17:26:38
上传
评论
收藏 4.29MB PDF 举报
温馨提示
试读
10页
Hierarchical Aggregation for 3D Instance Segmentation
资源推荐
资源详情
资源评论
Hierarchical Aggregation for 3D Instance Segmentation
Shaoyu Chen
1
Jiemin Fang
2,1
Qian Zhang
3
Wenyu Liu
1
Xinggang Wang
1†
1
School of EIC, Huazhong University of Science & Technology
2
Institute of AI, Huazhong University of Science & Technology
3
Horizon Robotics
{shaoyuchen,jaminfong,liuwy,xgwang}@hust.edu.cn {qian01.zhang}@horizon.ai
Abstract
Instance segmentation on point clouds is a fundamen-
tal task in 3D scene perception. In this work, we propose
a concise clustering-based framework named HAIS, which
makes full use of spatial relation of points and point sets.
Considering clustering-based methods may result in over-
segmentation or under-segmentation, we introduce the hi-
erarchical aggregation to progressively generate instance
proposals, i.e., point aggregation for preliminarily cluster-
ing points to sets and set aggregation for generating com-
plete instances from sets. Once the complete 3D instances
are obtained, a sub-network of intra-instance prediction is
adopted for noisy points filtering and mask quality scoring.
HAIS is fast (only 410ms per frame on Titan X)) and does
not require non-maximum suppression. It ranks 1st on the
ScanNet v2 benchmark
1
, achieving the highest 69.9% AP
50
and surpassing previous state-of-the-art (SOTA) methods by
a large margin. Besides, the SOTA results on the S3DIS
dataset validate the good generalization ability. Code is
available at https://github.com/hustvl/HAIS.
1. Introduction
With the rapid development and popularization of com-
modity 3D sensors (Kinect, RealSense, Velodyne laser
scanner, etc.), 3D scene understanding has become a hot
research topic in the field of computer vision. Instance
segmentation on point cloud, as the basic perception task
of 3D scene understanding, is the technical foundation of
a wide range of real-life applications, e.g., robotics, aug-
mented/virtual reality, and autonomous driving.
Instance segmentation on 2D images has been exhaus-
tively studied in the past few years [7, 24, 16, 27, 18, 4, 5,
1
http://kaldir.vc.in.tum.de/scannet_benchmark/
semantic_instance_3d
†
Xinggang Wang is the corresponding author.
point cloud
ground truth
w/ hierarchical aggrega�on
w/o hierarchical aggrega�on
Figure 1. An input point cloud, ground truth instance masks and
3D instance prediction results without & with hierarchical aggre-
gation. As shown in the key region circled with red, for objects
with large sizes and fragmentary point clouds, the predictions are
easy to be over-segmented. The proposed hierarchical aggregation
combines incomplete instances with fragments to form complete
instance predictions.
11]. Top-down methods dominate 2D instance segmenta-
tion. They first generate instance-level proposals and then
predict the mask for each proposal. Though existing 2D
instance segmentation methods can be directly extended to
3D scenes, most existing 3D methods adopt a totally differ-
ent bottom-up pipeline [42, 41, 22, 15, 20], which generates
instances through clustering.
However, directly clustering a point cloud into multiple
instances is very difficult for the following reasons: (1) A
point cloud usually contains a large number of points; (2)
The number of instances in a point cloud has large varia-
tions for different 3D scenes; (3) The sizes of instances vary
significantly; (4) Each point has a very weak feature, i.e.,
3D coordinate and color. The semantic gap between point
and instance identity is huge. Thus, over-segmentation or
under-segmentation are common problems and are prone to
15467
资源评论
易小侠
- 粉丝: 6514
- 资源: 9万+
上传资源 快速赚钱
- 我的内容管理 展开
- 我的资源 快来上传第一个资源
- 我的收益 登录查看自己的收益
- 我的积分 登录查看自己的积分
- 我的C币 登录后查看C币余额
- 我的收藏
- 我的下载
- 下载帮助
安全验证
文档复制为VIP权益,开通VIP直接复制
信息提交成功