[{"data":1,"prerenderedAt":296},["ShallowReactive",2],{"content-query-9cM3Pz1Hv2":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"cover":11,"type":12,"body":13,"_type":290,"_id":291,"_source":292,"_file":293,"_stem":294,"_extension":295},"/technology-blogs/zh/3016","zh",false,"","MindSpore AI科学计算系列 | 三维形状表征","作者：于璠 来源：知乎","2024-03-05","https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2024/03/08/07b8a7ec813f458f9dd2fb5b8e6e35fc.png","technology-blogs",{"type":14,"children":15,"toc":285},"root",[16,24,43,51,56,61,69,74,79,86,91,96,103,108,113,120,125,130,137,142,147,154,159,166,171,176,186,196,205,218,229,234,242,253,263,274],{"type":17,"tag":18,"props":19,"children":21},"element","h1",{"id":20},"mindspore-ai科学计算系列-三维形状表征",[22],{"type":23,"value":8},"text",{"type":17,"tag":25,"props":26,"children":27},"p",{},[28,30,36,38],{"type":23,"value":29},"**作者：**",{"type":17,"tag":31,"props":32,"children":33},"strong",{},[34],{"type":23,"value":35},"于璠",{"type":23,"value":37}," ",{"type":17,"tag":31,"props":39,"children":40},{},[41],{"type":23,"value":42},"来源：知乎",{"type":17,"tag":25,"props":44,"children":45},{},[46],{"type":17,"tag":31,"props":47,"children":48},{},[49],{"type":23,"value":50},"背景",{"type":17,"tag":25,"props":52,"children":53},{},[54],{"type":23,"value":55},"近年来，AI在二维图像的识别、分割、生成等任务中取得了丰硕的成果，然而，对于我们生活的三维世界，三维形状则更为普遍，因此，在Midjourney等图像生成活跃发展的今天，有必要考虑三维形状的表征和生成，本文将对三维形状的表示做一个简单的调研。",{"type":17,"tag":25,"props":57,"children":58},{},[59],{"type":23,"value":60},"我们该如何高效、精确的表达一个三维物体？以下图的Stanford Bunny为例：",{"type":17,"tag":25,"props":62,"children":63},{},[64],{"type":17,"tag":65,"props":66,"children":68},"img",{"alt":7,"src":67},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2024/03/08/2a345ac8c853411f9294707a2bebf310.png",[],{"type":17,"tag":25,"props":70,"children":71},{},[72],{"type":23,"value":73},"图1. Stanford Bunny",{"type":17,"tag":25,"props":75,"children":76},{},[77],{"type":23,"value":78},"一个很朴素的想法，就是参考二维的图像中像素点的概念，将三维外形表示为一个个三维像素，这样在三维空间中，就可以像二维图像一样，采用卷积来处理，但是缺点也很明显，这种表达的效率非常低，我们主要关注的是物体表面，但是我们需要同时处理物体的内部和外部的信息，这些信息是大量冗余的，尤其是我们需要对物体进行非常精细化描述的时候，网格会打得很细，导致占据的存储空间大大增加。",{"type":17,"tag":25,"props":80,"children":81},{},[82],{"type":17,"tag":65,"props":83,"children":85},{"alt":7,"src":84},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2024/03/08/4bfcd40b1e52464fbd7df2a1c0da42d5.png",[],{"type":17,"tag":25,"props":87,"children":88},{},[89],{"type":23,"value":90},"图2. 三位像素点表示的Stanford Bunny",{"type":17,"tag":25,"props":92,"children":93},{},[94],{"type":23,"value":95},"那么换一种想法，我们也可以用“拍照片”的方式描述一个三维形状，这样图像的处理方法、数据集就可以直接应用了，当然这种方式的缺点也很明显，照片拍摄的“死角”是很难避免的，导致信息的丢失。",{"type":17,"tag":25,"props":97,"children":98},{},[99],{"type":17,"tag":65,"props":100,"children":102},{"alt":7,"src":101},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2024/03/08/701bf94216be44038c755ba8361cc57a.png",[],{"type":17,"tag":25,"props":104,"children":105},{},[106],{"type":23,"value":107},"图3. “拍照片”方式下的Stanford Bunny",{"type":17,"tag":25,"props":109,"children":110},{},[111],{"type":23,"value":112},"更进一步，我们用网格的方式也可以表达一个三维形状，在表面通过记录节点坐标和节点间的连接信息，形成一个表形的结构。这种表达方式无疑是十分高效的，在相对平缓的平面上，我们可以少布一些点，在外形变化剧烈的地方，可以多放一些，因此，这种方式在计算图像学里应用十分广泛，在工程上的CAE分析中，也常用于表征固体的变形、流体的表面压力分布等等。在AI领域中，网格的数据结构也与图的概念相符，因此GNN，MeshCNN等方法可以进行针对性的处理。",{"type":17,"tag":25,"props":114,"children":115},{},[116],{"type":17,"tag":65,"props":117,"children":119},{"alt":7,"src":118},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2024/03/08/a3d8f942662c446cbf29ded1d5a6610d.png",[],{"type":17,"tag":25,"props":121,"children":122},{},[123],{"type":23,"value":124},"图4. 网格表示方式下的Stanford Bunny",{"type":17,"tag":25,"props":126,"children":127},{},[128],{"type":23,"value":129},"三维网格需要同时记录网格点的坐标和连接信息，由于连接信息常常十分复杂，那么能不能只保留表面点的坐标呢？那就是点云了，点云是物体表面无序点的集合，可以同时表示物体局部和整体的形状信息，另外，点云的数据非常容易获取，可以通过扫描、采样等方式直接得到。当前已经有PointNet、PointNet++、PointTransformer等模型用于点云的分类、分割等任务。",{"type":17,"tag":25,"props":131,"children":132},{},[133],{"type":17,"tag":65,"props":134,"children":136},{"alt":7,"src":135},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2024/03/08/baad7f44ede8483c9783d8fd2179cbbb.png",[],{"type":17,"tag":25,"props":138,"children":139},{},[140],{"type":23,"value":141},"图5. 点云表示下的Stanford Bunny",{"type":17,"tag":25,"props":143,"children":144},{},[145],{"type":23,"value":146},"最后，将神经压缩的想法应用在三维外形上，可以形成三维形状的隐式表达。以NeRF的方式为例，神经网络的输入为观测的位置和角度，输出则为该视角上的RGB和体密度值（原文的解释为光线在改点终止的概率），这样我们只需要保证每个视角下的照片和NeRF的渲染结果相一致，那么就可以对三维外形进行全面描述。对于每一个三维外形，我们都可以训一个NeRF，这样NeRF的参数就可以作为一个三维外形的隐式表征。由于其与观察视角的相符性，因此NeRF的渲染结果可以在手机/网页浏览器/VR/AR上自由操作，然而，NeRF的对于训练资源的需求较高，训练耗时长，尤其是复杂光照等条件中。",{"type":17,"tag":25,"props":148,"children":149},{},[150],{"type":17,"tag":65,"props":151,"children":153},{"alt":7,"src":152},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2024/03/08/a8bb9bb135684060b54525a7dd5435ed.png",[],{"type":17,"tag":25,"props":155,"children":156},{},[157],{"type":23,"value":158},"图6. NeRF(Neural Radiance Fields)",{"type":17,"tag":25,"props":160,"children":161},{},[162],{"type":17,"tag":65,"props":163,"children":165},{"alt":7,"src":164},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2024/03/08/26105cd355fe4cf5b0d9473bd6476ffc.png",[],{"type":17,"tag":25,"props":167,"children":168},{},[169],{"type":23,"value":170},"图7. NeRF表示下的Stanford Bunny",{"type":17,"tag":25,"props":172,"children":173},{},[174],{"type":23,"value":175},"最后分享两个常用的三维外形数据集：",{"type":17,"tag":25,"props":177,"children":178},{},[179,184],{"type":17,"tag":31,"props":180,"children":181},{},[182],{"type":23,"value":183},"ModelNet40",{"type":23,"value":185},"[1]**：**ModelNet40数据集包含合成对象点云。作为使用最广泛的点云分析基准，ModelNet40因其类别丰富、形状清晰、数据集构造良好等特点而广受欢迎。最初的ModelNet40由cad生成的40个类别(如飞机、汽车、工厂、灯具)的12,311个网格组成，其中9,843个用于训练，其余2,468个用于测试。从网格表面均匀采样相应的点云数据点，然后移动到原点进行了归一化处理。",{"type":17,"tag":25,"props":187,"children":188},{},[189,194],{"type":17,"tag":31,"props":190,"children":191},{},[192],{"type":23,"value":193},"ShapeNet",{"type":23,"value":195},"[2]**：**ShapeNet是由斯坦福大学、普林斯顿大学和美国芝加哥丰田技术研究所的研究人员开发的大型3D CAD模型存储库。该存储库包含超过300万个模型，其中22万个模型被分类为3135个类，使用WordNet上下名关系排列。ShapeNet Parts子集包含31,693个网格，分为16个常见对象类(如桌子、椅子、平面等)。每个形状的基础真理包含2-5个部分(总共50个部分类)。",{"type":17,"tag":197,"props":198,"children":200},"h2",{"id":199},"参考文献",[201],{"type":17,"tag":31,"props":202,"children":203},{},[204],{"type":23,"value":199},{"type":17,"tag":25,"props":206,"children":207},{},[208,210],{"type":23,"value":209},"[1] ",{"type":17,"tag":211,"props":212,"children":216},"a",{"href":213,"rel":214},"https://modelnet.cs.princeton.edu",[215],"nofollow",[217],{"type":23,"value":213},{"type":17,"tag":25,"props":219,"children":220},{},[221,223],{"type":23,"value":222},"[2] ",{"type":17,"tag":211,"props":224,"children":227},{"href":225,"rel":226},"https://shapenet.org",[215],[228],{"type":23,"value":225},{"type":17,"tag":25,"props":230,"children":231},{},[232],{"type":23,"value":233},"往期回顾",{"type":17,"tag":18,"props":235,"children":236},{"id":7},[237],{"type":17,"tag":211,"props":238,"children":241},{"href":239,"rel":240},"http://mp.weixin.qq.com/s?__biz=MzkxMTM2MjMzNg==&mid=2247615331&idx=1&sn=b5f6a74a14168a26b03f4d3c854bb5dc&chksm=c11e132cf6699a3a583f4b923f5060f0718799d260d67c90c134b14ff9ec3ab387c70d42766a&scene=21#wechat_redirect",[215],[],{"type":17,"tag":18,"props":243,"children":245},{"id":244},"mindspore-ai科学计算系列-生成扩散模型diffdock显著提升分子对接准确率",[246],{"type":17,"tag":211,"props":247,"children":250},{"href":248,"rel":249},"http://mp.weixin.qq.com/s?__biz=MzkxMTM2MjMzNg==&mid=2247615942&idx=1&sn=769392387b771f869bb20a69ce77ae69&chksm=c11e1589f6699c9f2d106880746621abad8d3bd3009669b11d0754f39d568e98e6e6626e757d&scene=21#wechat_redirect",[215],[251],{"type":23,"value":252},"MindSpore AI科学计算系列 | 生成扩散模型DiffDock显著提升分子对接准确率",{"type":17,"tag":18,"props":254,"children":256},{"id":255},"mindspore-ai科学计算系列-meshgpt显著提升三维几何表示的质量",[257],{"type":17,"tag":211,"props":258,"children":260},{"href":239,"rel":259},[215],[261],{"type":23,"value":262},"MindSpore AI科学计算系列 | MeshGPT显著提升三维几何表示的质量",{"type":17,"tag":18,"props":264,"children":266},{"id":265},"mindspore-ai科学计算系列-metnet3融合稀疏站点数据实现提前24h天气预报",[267],{"type":17,"tag":211,"props":268,"children":271},{"href":269,"rel":270},"http://mp.weixin.qq.com/s?__biz=MzkxMTM2MjMzNg==&mid=2247614540&idx=1&sn=08b091b4e629243defcdf97270234d88&chksm=c11e1003f6699915db3166681b1cb30c321964d71c0b1b362ba62d7ac3f1f05f2586b396be25&scene=21#wechat_redirect",[215],[272],{"type":23,"value":273},"MindSpore AI科学计算系列 | MetNet3融合稀疏站点数据，实现提前24h天气预报",{"type":17,"tag":18,"props":275,"children":277},{"id":276},"mindspore-ai科学计算系列-化学深度学习模型chemgpt的性能评估公式拟合",[278],{"type":17,"tag":211,"props":279,"children":282},{"href":280,"rel":281},"http://mp.weixin.qq.com/s?__biz=MzkxMTM2MjMzNg==&mid=2247614481&idx=1&sn=e45cb91a0dc5f51db7e80761d639a474&chksm=c11e105ef66999488545e230a32ef8c771768c8b488024d5b039bba9ed12974339f795cbe42e&scene=21#wechat_redirect",[215],[283],{"type":23,"value":284},"MindSpore AI科学计算系列 | 化学深度学习模型ChemGPT的性能评估公式拟合",{"title":7,"searchDepth":286,"depth":286,"links":287},4,[288],{"id":199,"depth":289,"text":199},2,"markdown","content:technology-blogs:zh:3016.md","content","technology-blogs/zh/3016.md","technology-blogs/zh/3016","md",1776506125356]