论文笔记：partNet

PartNet: A Large-scale Benchmark for Fine-grained and Hierarchical Part-level 3D Object Understanding

采用树状分层的方式分割物体，分成了coarse,middle,fine三个等级。

1、Contribution

• large-scale: 573,585 finegrained part annotations for 26,671 shapes across 24 object categories
• propose three part-level object understanding tasks: fine-grained semantic segmentation, hierarchical semantic segmentation, and instance segmentation.
• benchmark four algorithm for semantic segmentation and three baseline methods for hierarchical segmentation
• a novel method for part instance segmentation
  

2、Data Annotation

2.1 Expert-Defined Part Hierarchy

The criteria that use to guide template design;

• well defined
• consistent
• conpact
• hierarchical
• atomic
• complete

2.2 Annotation Interface

3.PartNet Dataset

26,671 shapes with 573,585 part instances from 24 object categories
Most of the shapes and object categories are from ShapeNetCore

4.Tasks and Benchmarks

Data Preparation

只考虑可以完全由形状几何图形决定的零件,像橱柜上的玻璃，微波炉上的按钮就不考虑。每个CAD模型用FPS采样10000个点。

4.1 Fine-grained Semantic Segmentation

Benchmark Algorithms

fine-grained PartNet segmentation: PointNet, PointNet++, SpiderCNN, PointCNN

Evaluation and Results

评价分为了细粒度，中粒度，粗粒度三个等级
计算每个类里面的每个部分IoU

使用mean IoU
evaluate the algorithms at three segmentation levels for each object category: coarse-, middle- and fine-grained
use mean Intersection-over-Union(mIoU)， the per-part-category IoUs

5.2 Hierarchical Semantic Segmentation

任务是预测semantic part labels
整个形状中粗粒度和细粒度部分概念的semantic part label
一个重要的问题是怎样利用 rich part relationships on the given shape templates in the learning procedure

Benchmark Algorithms

hierarchical segmentation: bottom-up, top-down and ensemble
bottom-up:在训练过程中只考虑叶节点部分，并将子节点对父节点的预测按自底向上推理的层次结构定义进行分组。
top-down: 学习树上所有部件语义标签的多重标记任务，并通过对较粗级别节点和较低级别节点进行分类来进行自顶向下的推理。
ensemble:在多个层次上进行训练细分。通过计算所有根到叶路径树的平均对数似然得分进行联合推断
使用PointNet ++

Evaluation and Results

对每个类别的part计算mIoU，然后对所有的节点计算均值。三个方法结果都差不多，ensemble更胜一筹。

5.3 Instance Segmentation

目标：检测shape里面的每个部分并且分割出来

Benchmark Algorithms

使用**PointNet++**作为backbone，对每个点预测语义分割结果并预测实例分割maskKaTeX parse error: Expected '}', got 'EOF' at end of input: …N|i = 1,2,...,K，对于在gt没有semantic label的点,记为${\hat{y}}_{other}$，然后使用softmax激活层，满足$${\hat{y}}_1+{\hat{y}}_2+...+{\hat{y}}_K+{\hat{y}}_{other}=1$$.训练网络的时候用上了Hungarian algorithm 寻找gt和predict的匹配$M:i\to M(i)|i=1,2,..,T$.然后回归预测的${\hat{y}}_{M(t)}$与匹配gt$y_t$。

Evaluation and Results.

mAP (mean Average Precision)