Large-Scale 3D Shape Reconstruction and Segmentation from ShapeNet Core55
Li Yi1 Lin Shao1 Manolis Savva2 Haibin Huang3 Yang Zhou3 Qirui Wang4 Benjamin Graham5 Martin Engelcke6 Roman Klokov7 Victor Lempitsky7 Yuan Gan8 Pengyu Wang8 Kun Liu8 Fenggen Yu8 Panpan Shui8 Bingyang Hu8 Yan Zhang8 Yangyan Li9 Rui Bu9 Mingchao Sun9 Wei Wu9 Minki Jeong10 Jaehoon Choi10 Changick Kim10 Angom Geetchandra11 Narasimha Murthy11 Bhargava Ramu11 Bharadwaj Manda11 M Ramanathan11 Gautam Kumar13 Preetham P13 Siddharth Srivastava13 Swati Bhugra13 Brejesh Lall13 Christian Häne14 Shubham Tulsiani14 Jitendra Malik14 Jared Lafer15 Ramsey Jones15 Siyuan Li16 Jie Lu16 Shi Jin16 Jingyi Yu16 Qixing Huang17 Evangelos Kalogerakis3 Silvio Savarese1 Pat Hanrahan1 Thomas Funkhouser2 Hao Su12 Leonidas Guibas1 1Stanford University 2Princeton University 3University of Massachusetts–Amherst 4Tsinghua University 5Facebook AI Research 6University of Oxford 7Skolkovo Institute of Science and Technology 8Nanjing University 9Shandong University 10Korea Advanced Institute of Science and Technology 11Indian Institute of Technology Madras 12UC San Diego 13Indian Institute of Technology Delhi 14UC Berkeley 15Imbellus 16ShanghaiTech University 17University of Texas, Austin
Abstract
We introduce a large-scale 3D shape understanding benchmark using data and annotation from ShapeNet 3D object database. The benchmark consists of two tasks: part-level segmentation of 3D shapes and 3D reconstruction from single view images. Ten teams have participated in the challenge and the best performing teams have outperformed state-of-the-art approaches on both tasks. A few novel deep learning architectures have been proposed on various 3D representations on both tasks. We report the techniques used by each team and the corresponding performances. In addition, we summarize the major discoveries from the reported results and possible trends for the future work in the field.
中文速览
三维形状的理解与重建长期缺乏公认的大规模评测基准，ShapeNet团队因此推出了这一挑战赛，涵盖两项任务：基于点云的三维形状零件级语义分割（part-level segmentation），以及从单张图像重建三维形状（single-view 3D reconstruction）。十支参赛队伍全部采用深度学习方法，并在体素（voxel）、点云（point cloud）等多种三维表示形式上提出了各具特色的网络架构，其中点云表示尤为受到青睐，涌现出稀疏卷积网络、密集连接PointNet、PointCNN等新颖设计。最终，各队表现均超越了赛事组织方实现的现有基线，但有趣的是，在重建任务中，以IoU为指标和以Chamfer距离（Chamfer Distance）为指标时，夺冠的是两种不同方法，表明评测指标的选取本身仍是一个值得深入探讨的开放问题。这一基准的建立为三维视觉领域提供了统一的评测平台，有助于整合计算机图形学、计算机视觉和机器学习社区的研究合力，推动后续工作在可比条件下持续演进。
原文 arXiv:1710.06104;中英对照 + 大白话阅读 https://aha.fim.ai/paper/1710.06104v2