表面厂家
免费服务热线

Free service

hotline

010-00000000
表面厂家
热门搜索:
行业资讯
当前位置:首页 > 行业资讯

CCCV2017讲习班笔记基于图像的大规模场景三维重建下_[#第一枪]

发布时间:2021-06-07 13:10:37 阅读: 来源:表面厂家

AI科技评论按,本文作者究竟灰,本文首发于知乎,雷锋网 AI科技评论获其授权转载。

雷锋网注:本文为下篇,内容为第三章:稠密重建和第四章:稠密重建。第一章和第二章参见CCCV2017讲习班笔记-基于图像的大规模场景三维重建(上)

<img data-rawheight="495" src="https://static.leiphone.com/uploads/new/article/pic/201711/1db4921ba9006fb90f6d3f10ec393c1a.jpg" data-rawwidth="702" class="origin_image zh-lightbox-thumb" width="702" data-original="https://pic3.zhimg.com/v2-86b6152d7c3a7f979254a5c8e073b6de_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/1db4921ba9006fb90f6d3f10ec393c1a.jpg"/>

3.稠密匹配

稠密匹配是MVS.基本思路是

<img data-rawheight="575" src="https://static.leiphone.com/uploads/new/article/pic/201711/b86045e10168e66d743fd92b986337dc.jpg" data-rawwidth="1123" class="origin_image zh-lightbox-thumb" width="1123" data-original="https://pic3.zhimg.com/v2-556e40393ecddac7789ba5169e74e942_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/b86045e10168e66d743fd92b986337dc.jpg"/>

两视图的一致性。

<img data-rawheight="679" src="https://static.leiphone.com/uploads/new/article/pic/201711/f6396ec552c37f46a475828800cdaba6.jpg" data-rawwidth="897" class="origin_image zh-lightbox-thumb" width="897" data-original="https://pic3.zhimg.com/v2-bf3e00c0f206da76b9b31af71b997bf2_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/f6396ec552c37f46a475828800cdaba6.jpg"/>

一致性度量主要包括三个:

SSD(Sum of Squared Differences):平常差的和

SAD(Sum of Absolute Differences):绝对值差的和

NCC(Normalized Cross Correlation):归一化的交叉关系

<img data-rawheight="585" src="https://static.leiphone.com/uploads/new/article/pic/201711/9741ebacf34e7d7be4f61f22cc9c1671.jpg" data-rawwidth="956" class="origin_image zh-lightbox-thumb" width="956" data-original="https://pic3.zhimg.com/v2-2498d412f51509f9efe6bbafa1f14716_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/9741ebacf34e7d7be4f61f22cc9c1671.jpg"/>

多视图图像一致性需要考虑相机的可视性问题。

但是:相机可视性需要场景结构、场景结构需要相机可视性

<img data-rawheight="423" src="https://static.leiphone.com/uploads/new/article/pic/201711/81d65cbf7ae750d46c8ae7660f587341.jpg" data-rawwidth="793" class="origin_image zh-lightbox-thumb" width="793" data-original="https://pic3.zhimg.com/v2-84472f01ed66169f5bbafb798b233ea6_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/81d65cbf7ae750d46c8ae7660f587341.jpg"/>

MVS算法主要分为三种:

基于体素的方法:Voxel based MVS

基于点云扩散的方法:Feature point growing based MVS

基于深度图融合的方法:Depth-map merging based MVS

<img data-rawheight="513" src="https://static.leiphone.com/uploads/new/article/pic/201711/0378cff880de0c45a9acb9313d555b54.jpg" data-rawwidth="744" class="origin_image zh-lightbox-thumb" width="744" data-original="https://pic3.zhimg.com/v2-8851768c37ca6a3afa70a6ecb21a7652_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/0378cff880de0c45a9acb9313d555b54.jpg"/>

基于体素的方法

体素的表达,并且说明其MVS等价于一个3D空间Voxel的标记问题。

<img data-rawheight="632" src="https://static.leiphone.com/uploads/new/article/pic/201711/cae081ecab5cbdf65a5fbd2ee0fdfd8b.jpg" data-rawwidth="890" class="origin_image zh-lightbox-thumb" width="890" data-original="https://pic1.zhimg.com/v2-021dcde75f256c84bf6e78046112e38c_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/cae081ecab5cbdf65a5fbd2ee0fdfd8b.jpg"/>

优化方法:用马尔科夫随机场优化。

<img data-rawheight="651" src="https://static.leiphone.com/uploads/new/article/pic/201711/5aa45c03e05663104de5ca034e6a7021.jpg" data-rawwidth="1020" class="origin_image zh-lightbox-thumb" width="1020" data-original="https://pic4.zhimg.com/v2-1a0d829f145345b2785100a6716e6b8f_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/5aa45c03e05663104de5ca034e6a7021.jpg"/>

离散空间的Labeling是典型的MRF优化问题。其中的两项分别是一致性项和气球膨胀。

一致性项表达两点一致。气球膨胀表达的是强制倾向于把点分成内点。因为如果不加气球膨胀,一致性项会把点都分成外点,所以要加一个反向的力量。

<img data-rawheight="637" src="https://static.leiphone.com/uploads/new/article/pic/201711/31ae6999c84a84844173e98584bec256.jpg" data-rawwidth="966" class="origin_image zh-lightbox-thumb" width="966" data-original="https://pic1.zhimg.com/v2-716fcd006af1d3e00ccdb7086add6ad0_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/31ae6999c84a84844173e98584bec256.jpg"/>

两视图一致性计算:

<img data-rawheight="734" src="https://static.leiphone.com/uploads/new/article/pic/201711/7a719664e7eda57551918d864f32a309.jpg" data-rawwidth="1023" class="origin_image zh-lightbox-thumb" width="1023" data-original="https://pic4.zhimg.com/v2-d9b038cfe8a30d341d0a758d687f5467_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/7a719664e7eda57551918d864f32a309.jpg"/>

其中如何鲁棒投票寻找局部极值集中的点很重要。

<img data-rawheight="678" src="https://static.leiphone.com/uploads/new/article/pic/201711/e6be0b27123b4fd78f3fe0c0b78d90e7.jpg" data-rawwidth="1061" class="origin_image zh-lightbox-thumb" width="1061" data-original="https://pic3.zhimg.com/v2-826782c551e579ddf2d1e0eedede01fa_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/e6be0b27123b4fd78f3fe0c0b78d90e7.jpg"/>

MRF优化问题求解:Graph-cuts

<img data-rawheight="524" src="https://static.leiphone.com/uploads/new/article/pic/201711/79dc6a690712b8205ff5c3f320b780d6.jpg" data-rawwidth="735" class="origin_image zh-lightbox-thumb" width="735" data-original="https://pic4.zhimg.com/v2-acbe4226e76d834c4e0c1163e726fa07_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/79dc6a690712b8205ff5c3f320b780d6.jpg"/>

重建结果:

<img data-rawheight="503" src="https://static.leiphone.com/uploads/new/article/pic/201711/b8a3bbb858b011f76cbbb62d89a9287e.jpg" data-rawwidth="754" class="origin_image zh-lightbox-thumb" width="754" data-original="https://pic1.zhimg.com/v2-ba070a9a7a4fa9dab1c80b9bdd882f90_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/b8a3bbb858b011f76cbbb62d89a9287e.jpg"/>

体素问题是占内存,即使很小的体素也要很大内存。于是提出以下方法,主要思路是自适应多分辨率网格,在物体表面高分辨率、其他区域低分辨率。

<img data-rawheight="606" src="https://static.leiphone.com/uploads/new/article/pic/201711/d5bae9ed843162657bb8fcdc9d9744ec.jpg" data-rawwidth="981" class="origin_image zh-lightbox-thumb" width="981" data-original="https://pic3.zhimg.com/v2-469b02db65d1f9656ddca54731166bce_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/d5bae9ed843162657bb8fcdc9d9744ec.jpg"/>

基于体素方法MVS的并行分布Graph-cuts

<img data-rawheight="713" src="https://static.leiphone.com/uploads/new/article/pic/201711/585ea02328fd76f9385181a6fdb3729a.jpg" data-rawwidth="954" class="origin_image zh-lightbox-thumb" width="954" data-original="https://pic2.zhimg.com/v2-287b2bd6100d329dd0fa923aab31ee0d_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/585ea02328fd76f9385181a6fdb3729a.jpg"/>

基于体素方法的优缺点:

<img data-rawheight="637" src="https://static.leiphone.com/uploads/new/article/pic/201711/d454c52bb7a3cd66ec584ad98b9e0c4b.jpg" data-rawwidth="870" class="origin_image zh-lightbox-thumb" width="870" data-original="https://pic3.zhimg.com/v2-f077de0706f800f0e0076b42188c84d6_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/d454c52bb7a3cd66ec584ad98b9e0c4b.jpg"/>

基于特征点扩散的MVS

方法顾名思义。

<img data-rawheight="629" src="https://static.leiphone.com/uploads/new/article/pic/201711/dd405995c8542c88d75611ac6a629577.jpg" data-rawwidth="878" class="origin_image zh-lightbox-thumb" width="878" data-original="https://pic2.zhimg.com/v2-e9b387d7bd9041c2591856437893784d_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/dd405995c8542c88d75611ac6a629577.jpg"/>

讲了3D点的Patch形式表达。patch在图像上有投影。

<img data-rawheight="588" src="https://static.leiphone.com/uploads/new/article/pic/201711/06f9020ff2095614a6276085c1e3585e.jpg" data-rawwidth="861" class="origin_image zh-lightbox-thumb" width="861" data-original="https://pic1.zhimg.com/v2-e2d501279ae366c84d770141c68a51d8_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/06f9020ff2095614a6276085c1e3585e.jpg"/>

步骤:

生成初始点云:检测Harris与DoG,其中Harris偏向检测外侧的角点,而DoG偏向于检测内部纹理丰富的点

点云扩散:3D点投影到图像,并向投影点周围区域扩散

点云过滤:去除深度值不一致且一致性较低的点,意思是如果扩散的点云在其他图特征点的点云前面了,通过比较各自的一致性来剔除;如果扩散点云跑到后边去了,也比较一致性。这样就能去除深度值不一致且一致性较低的点了。

<img data-rawheight="666" src="https://static.leiphone.com/uploads/new/article/pic/201711/feff97af9dfbd5c6197c59413ba6f0fd.jpg" data-rawwidth="885" class="origin_image zh-lightbox-thumb" width="885" data-original="https://pic1.zhimg.com/v2-821c4b64769f3697f3b7dc4cfb45d92c_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/feff97af9dfbd5c6197c59413ba6f0fd.jpg"/><img data-rawheight="586" src="https://static.leiphone.com/uploads/new/article/pic/201711/2ffa8fc5f924624bb392252a9edb059a.jpg" data-rawwidth="792" class="origin_image zh-lightbox-thumb" width="792" data-original="https://pic4.zhimg.com/v2-5a51d4582d32397f903730c90917db87_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/2ffa8fc5f924624bb392252a9edb059a.jpg"/><img data-rawheight="515" src="https://static.leiphone.com/uploads/new/article/pic/201711/74b1d4df8c2cd6584c1a884f6af5dce8.jpg" data-rawwidth="868" class="origin_image zh-lightbox-thumb" width="868" data-original="https://pic2.zhimg.com/v2-8a6192136537e7df03121173fb7cee0d_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/74b1d4df8c2cd6584c1a884f6af5dce8.jpg"/>

结果:

<img data-rawheight="623" src="https://static.leiphone.com/uploads/new/article/pic/201711/985f2516b0e69d2b6a03850537a74a95.jpg" data-rawwidth="955" class="origin_image zh-lightbox-thumb" width="955" data-original="https://pic4.zhimg.com/v2-d98a371ebba0ce2649f715ea89b92c7b_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/985f2516b0e69d2b6a03850537a74a95.jpg"/>

优缺点:

<img data-rawheight="482" src="https://static.leiphone.com/uploads/new/article/pic/201711/06038cfbb7f1ec41a5b11c91a21c68fb.jpg" data-rawwidth="1032" class="origin_image zh-lightbox-thumb" width="1032" data-original="https://pic4.zhimg.com/v2-6abfc45975a14c07aff54b17877ac5c7_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/06038cfbb7f1ec41a5b11c91a21c68fb.jpg"/>

基于深度图融合的MVS

人的左右眼的立体视觉和深度图。

<img data-rawheight="644" src="https://static.leiphone.com/uploads/new/article/pic/201711/34facf9e4a679c0af33680a5147b4707.jpg" data-rawwidth="1046" class="origin_image zh-lightbox-thumb" width="1046" data-original="https://pic1.zhimg.com/v2-97d50ef5835c0e2115d05aaab96cf520_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/34facf9e4a679c0af33680a5147b4707.jpg"/>

转到CV

<img data-rawheight="654" src="https://static.leiphone.com/uploads/new/article/pic/201711/a2af7a8e846723e9567c499d707e338b.jpg" data-rawwidth="976" class="origin_image zh-lightbox-thumb" width="976" data-original="https://pic2.zhimg.com/v2-a42d67926daa818db7b579798f110729_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/a2af7a8e846723e9567c499d707e338b.jpg"/>

基于深度图融合的MVS方法步骤:

为每一幅图选择领域图像构成立体图像组:关键如何选择邻域图像组

计算每一幅图像的深度图:关键如何计算深度图

深度图融合

抽取物体表面

<img data-rawheight="450" src="https://static.leiphone.com/uploads/new/article/pic/201711/ffc1692055aca9932977e1d7e4bbee28.jpg" data-rawwidth="870" class="origin_image zh-lightbox-thumb" width="870" data-original="https://pic1.zhimg.com/v2-721f6baa5e381a6056444fd227e71358_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/ffc1692055aca9932977e1d7e4bbee28.jpg"/>

<img data-rawheight="518" src="https://static.leiphone.com/uploads/new/article/pic/201711/68d3778c77cc8f30aaa012781fefd6cd.jpg" data-rawwidth="833" class="origin_image zh-lightbox-thumb" width="833" data-original="https://pic4.zhimg.com/v2-976b5c039855fc6dfbc75adf2d9bf5a7_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/68d3778c77cc8f30aaa012781fefd6cd.jpg"/><img data-rawheight="560" src="https://static.leiphone.com/uploads/new/article/pic/201711/cc4899f03ad2053f71b66cc157d136dc.jpg" data-rawwidth="834" class="origin_image zh-lightbox-thumb" width="834" data-original="https://pic2.zhimg.com/v2-4b83963d6391992a0062e3051556fc45_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/cc4899f03ad2053f71b66cc157d136dc.jpg"/><img data-rawheight="664" src="https://static.leiphone.com/uploads/new/article/pic/201711/cd4aedcbbede7a38dc386ddc1b62b117.jpg" data-rawwidth="1020" class="origin_image zh-lightbox-thumb" width="1020" data-original="https://pic3.zhimg.com/v2-22ee67a8f0523d465c598d81fca7e502_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/cd4aedcbbede7a38dc386ddc1b62b117.jpg"/>

每一幅图中的深度图计算:

<img data-rawheight="596" src="https://static.leiphone.com/uploads/new/article/pic/201711/55d39face3ed135b28bf4eb591019e20.jpg" data-rawwidth="897" class="origin_image zh-lightbox-thumb" width="897" data-original="https://pic4.zhimg.com/v2-1d03e6e511ef8756308504b3ae4b0023_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/55d39face3ed135b28bf4eb591019e20.jpg"/>

聚合:对比了SAD聚合和Adaptive weight

<img data-rawheight="670" src="https://static.leiphone.com/uploads/new/article/pic/201711/61fcecf19663972e52200b07345729a6.jpg" data-rawwidth="975" class="origin_image zh-lightbox-thumb" width="975" data-original="https://pic3.zhimg.com/v2-aad399a0d0418efecb76c686d434387a_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/61fcecf19663972e52200b07345729a6.jpg"/><img data-rawheight="588" src="https://static.leiphone.com/uploads/new/article/pic/201711/b6387a9523c742585d1f5c4e81c52d19.jpg" data-rawwidth="882" class="origin_image zh-lightbox-thumb" width="882" data-original="https://pic3.zhimg.com/v2-d6cb2274403de1b19b168622608a2356_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/b6387a9523c742585d1f5c4e81c52d19.jpg"/><img data-rawheight="670" src="https://static.leiphone.com/uploads/new/article/pic/201711/aad06a7c7111fbf2e9c5304854af9b45.jpg" data-rawwidth="983" class="origin_image zh-lightbox-thumb" width="983" data-original="https://pic2.zhimg.com/v2-fa1028063f10bc8528e25c8be20c6999_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/aad06a7c7111fbf2e9c5304854af9b45.jpg"/>

对比:

<img data-rawheight="524" src="https://static.leiphone.com/uploads/new/article/pic/201711/b43ec42e51a7d50eb86d13fd7d93dac5.jpg" data-rawwidth="864" class="origin_image zh-lightbox-thumb" width="864" data-original="https://pic1.zhimg.com/v2-20f3111084f61fef88f1df61963d0f48_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/b43ec42e51a7d50eb86d13fd7d93dac5.jpg"/>

这里讲了Oriented plane方法,估计空间平面方向

<img data-rawheight="757" src="https://static.leiphone.com/uploads/new/article/pic/201711/d0420233bf51245b842eb4f98973498d.jpg" data-rawwidth="1044" class="origin_image zh-lightbox-thumb" width="1044" data-original="https://pic4.zhimg.com/v2-b115c0b37b53849500e84c1f24a2c5bf_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/d0420233bf51245b842eb4f98973498d.jpg"/>

空间平面方向估计PathMatch,相机坐标系下空间面片表达为d深度的一个自由度,n法向量的两个自由度。

<img data-rawheight="514" src="https://static.leiphone.com/uploads/new/article/pic/201711/0f75eca913aafabf2507286e567167f4.jpg" data-rawwidth="828" class="origin_image zh-lightbox-thumb" width="828" data-original="https://pic3.zhimg.com/v2-87f820f94fbdad6990af5d8e83e4c49e_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/0f75eca913aafabf2507286e567167f4.jpg"/>

接着讲了两视图PathMatch Stereo,先随机申城像素深度和法向量,然后传播。

主要用了随机的思想,检测领域点的深度和法向量,检测加了扰动之后的点,检测立体图像对对应点是否更好,检测前后帧同一位置是否更好。反复几次。

<img data-rawheight="488" src="https://static.leiphone.com/uploads/new/article/pic/201711/36703632790fe85746fc5d8fe830c541.jpg" data-rawwidth="799" class="origin_image zh-lightbox-thumb" width="799" data-original="https://pic4.zhimg.com/v2-ed4aefd674d12aad49c63a01bde600b3_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/36703632790fe85746fc5d8fe830c541.jpg"/>

这种方法是基于大数定律的。

<img data-rawheight="130" src="https://static.leiphone.com/uploads/new/article/pic/201711/95855ce5c1f293f25769501e2ecf19ba.jpg" data-rawwidth="473" class="origin_image zh-lightbox-thumb" width="473" data-original="https://pic2.zhimg.com/v2-ddb596b837cd3d1687e2441ff5ac6391_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/95855ce5c1f293f25769501e2ecf19ba.jpg"/>

多视图PathMatch MVS:

<img data-rawheight="583" src="https://static.leiphone.com/uploads/new/article/pic/201711/3824d1d0626decad02d2edb3f375852f.jpg" data-rawwidth="914" class="origin_image zh-lightbox-thumb" width="914" data-original="https://pic1.zhimg.com/v2-5d968e5fd721768afe3681209a08e560_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/3824d1d0626decad02d2edb3f375852f.jpg"/>

多视图PathMatch中领域图像组的选择:视线夹角、物距、覆盖度、分散度等。主要通过稀疏点云计算。领域图像组选择是一个NP-hard问题。

<img data-rawheight="652" src="https://static.leiphone.com/uploads/new/article/pic/201711/d4d081f73d71371f7721b95da044c9e8.jpg" data-rawwidth="860" class="origin_image zh-lightbox-thumb" width="860" data-original="https://pic2.zhimg.com/v2-bcfa850914f01a3895bf27307773d70d_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/d4d081f73d71371f7721b95da044c9e8.jpg"/><img data-rawheight="711" src="https://static.leiphone.com/uploads/new/article/pic/201711/03fd96510e226dd6dfdaf220ccb34c8c.jpg" data-rawwidth="993" class="origin_image zh-lightbox-thumb" width="993" data-original="https://pic4.zhimg.com/v2-8c5371cc274b03c49628a4b2f5cf9f5f_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/03fd96510e226dd6dfdaf220ccb34c8c.jpg"/>

逐像素点领域选择:

<img data-rawheight="479" src="https://static.leiphone.com/uploads/new/article/pic/201711/6e2ab9db02f568d9bfd60a11bf980179.jpg" data-rawwidth="727" class="origin_image zh-lightbox-thumb" width="727" data-original="https://pic2.zhimg.com/v2-2a7ee3b033fc54b5694460b0eda382ad_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/6e2ab9db02f568d9bfd60a11bf980179.jpg"/><img data-rawheight="423" src="https://static.leiphone.com/uploads/new/article/pic/201711/5086e2d70358db1124c2ea0e1efe2901.jpg" data-rawwidth="628" class="origin_image zh-lightbox-thumb" width="628" data-original="https://pic2.zhimg.com/v2-09333ae8928c12db8fa7835930009a7d_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/5086e2d70358db1124c2ea0e1efe2901.jpg"/><img data-rawheight="510" src="https://static.leiphone.com/uploads/new/article/pic/201711/7a03c5eef79a41ae8d4b9bb36d2248c4.jpg" data-rawwidth="850" class="origin_image zh-lightbox-thumb" width="850" data-original="https://pic2.zhimg.com/v2-1db064f165fc08e36cd3ea3d16b9c279_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/7a03c5eef79a41ae8d4b9bb36d2248c4.jpg"/>

通过EM算法来做逐像素点选择领域图像组(最大化后验概率)

<img data-rawheight="643" src="https://static.leiphone.com/uploads/new/article/pic/201711/9f569491e4d4357c34d09a65cd58ea90.jpg" data-rawwidth="847" class="origin_image zh-lightbox-thumb" width="847" data-original="https://pic2.zhimg.com/v2-50d4b7e6db12457e3628442092e50a49_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/9f569491e4d4357c34d09a65cd58ea90.jpg"/>

基于深度图融合的MVS优缺点:

<img data-rawheight="435" src="https://static.leiphone.com/uploads/new/article/pic/201711/21e6aad2ba23973f1b28c7d256f53881.jpg" data-rawwidth="753" class="origin_image zh-lightbox-thumb" width="753" data-original="https://pic1.zhimg.com/v2-4c8c5fc3018452c18473a30ea644e814_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/21e6aad2ba23973f1b28c7d256f53881.jpg"/>

稠密重建总结

<img data-rawheight="465" src="https://static.leiphone.com/uploads/new/article/pic/201711/a6c9fd01107e001e6d17e5bb9635daae.jpg" data-rawwidth="807" class="origin_image zh-lightbox-thumb" width="807" data-original="https://pic1.zhimg.com/v2-d188af740056493fd9a9cf80014fe154_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/a6c9fd01107e001e6d17e5bb9635daae.jpg"/>

4.资源

主要是算法、数据集和应用

<img data-rawheight="592" src="https://static.leiphone.com/uploads/new/article/pic/201711/166916bc177cedcc1235afb5331ab72e.jpg" data-rawwidth="963" class="origin_image zh-lightbox-thumb" width="963" data-original="https://pic1.zhimg.com/v2-08b7a6857ae0a34e38eedf8aa6869f80_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/166916bc177cedcc1235afb5331ab72e.jpg"/><img data-rawheight="429" src="https://static.leiphone.com/uploads/new/article/pic/201711/ed221dd979b7e39300e24ba2314add7d.jpg" data-rawwidth="653" class="origin_image zh-lightbox-thumb" width="653" data-original="https://pic2.zhimg.com/v2-5f623b01dffe42eb36427925a3cfc991_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/ed221dd979b7e39300e24ba2314add7d.jpg"/><img data-rawheight="637" src="https://static.leiphone.com/uploads/new/article/pic/201711/a14dee1869187de6825e9fad4bdb6dde.jpg" data-rawwidth="872" class="origin_image zh-lightbox-thumb" width="872" data-original="https://pic3.zhimg.com/v2-c575bf88efee4e3ecabb013ab846483e_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/a14dee1869187de6825e9fad4bdb6dde.jpg"/><img data-rawheight="636" src="https://static.leiphone.com/uploads/new/article/pic/201711/2f8fcd023efe1aa92974d1d5bef1ab22.jpg" data-rawwidth="1028" class="origin_image zh-lightbox-thumb" width="1028" data-original="https://pic2.zhimg.com/v2-eb542a9ade0659dbabc24c8860545589_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/2f8fcd023efe1aa92974d1d5bef1ab22.jpg"/><img data-rawheight="658" src="https://static.leiphone.com/uploads/new/article/pic/201711/c7857bd1e9d2e21775f2112ba2ae8c59.jpg" data-rawwidth="858" class="origin_image zh-lightbox-thumb" width="858" data-original="https://pic2.zhimg.com/v2-13b2ae9674aeedafe596bc9f657a6305_r.jpg" _src="https://static.leiphone.com/uploads/new/article/pic/201711/c7857bd1e9d2e21775f2112ba2ae8c59.jpg"/>

雷锋网版权文章,未经授权禁止转载。详情见转载须知。

迎宾踏板安装图片

垂钓用品套件批发

履带撒肥车批发

相关阅读