https://github.com/galmetzer/dipole-normal-prop
Tip revision: 0887b893b153f4ce6b09e4dd485f9b042f15b706 authored by Gal Metzer on 08 September 2021, 21:53:11 UTC
Update README.md
Update README.md
Tip revision: 0887b89
index.html
<!DOCTYPE html>
<head>
<meta charset="utf-8"/>
<title>Dipole Propagation</title>
<link rel="stylesheet" href="https://use.fontawesome.com/releases/v5.8.1/css/all.css"
integrity="sha384-50oBUHEmvpQ+1lW4y57PTFmhCaXp0ML5d60M1M7uH2+nqUivzIebhndOJK28anvf" crossorigin="anonymous">
<meta content="width=device-width, initial-scale=1" name="viewport"/>
<link href="min.css" rel="stylesheet" type="text/css"/>
<script src="https://ajax.googleapis.com/ajax/libs/webfont/1.6.26/webfont.js" type="text/javascript"></script>
<script type="text/javascript">
WebFont.load({google: {families: ["Lato:100,100italic,300,300italic,400,400italic,700,700italic,900,900italic", "Montserrat:100,100italic,200,200italic,300,300italic,400,400italic,500,500italic,600,600italic,700,700italic,800,800italic,900,900italic", "Ubuntu:300,300italic,400,400italic,500,500italic,700,700italic", "Changa One:400,400italic", "Open Sans:300,300italic,400,400italic,600,600italic,700,700italic,800,800italic", "Varela Round:400", "Bungee Shade:regular", "Roboto:300,regular,500"]}});</script>
<!--[if lt IE 9]>
<script src="https://cdnjs.cloudflare.com/ajax/libs/html5shiv/3.7.3/html5shiv.min.js"
type="text/javascript"></script><![endif]-->
<script type="text/javascript">!function (o, c) {
var n = c.documentElement, t = " w-mod-";
n.className += t + "js", ("ontouchstart" in o || o.DocumentTouch && c instanceof DocumentTouch) && (n.className += t + "touch")
}(window, document);</script>
<link href="images/thumb.png"
rel="shortcut icon" type="image/x-icon"/>
<style>
.wf-loading * {
opacity: 0;
}
</style>
</head>
<body>
<div class="row">
<div class="cell"><img src="images/s2021-simple-logo.png"></div>
</div>
<div class="section hero p2m">
<div class="container-2 p2m_header_v2 w-container"><h1 class="title">Orienting Point Clouds with Dipole Propagation</h1>
<h1 class="subheader"> TLDR; Leveraging the power of neural networks and a robust and stable dipole propagation technique </h1>
<div class="p2m_authors_list_single w-row">
<div class="w-col w-col-3 w-col-small-2 w-col-tiny-6">
<a class="authors" href="https://galmetzer.github.io/" target="_blank">Gal Metzer<sup>1</sup></a>
</div>
<div class="w-col w-col-3 w-col-small-4 w-col-tiny-6">
<a class="authors" href="https://www.cs.tau.ac.il/~hanocka/" target="_blank">Rana Hanocka<sup>1</sup></a>
</div>
<div class="w-col w-col-3 w-col-small-2 w-col-tiny-6">
<a class="authors" href="https://cims.nyu.edu/gcl/denis.html" target="_blank">Denis Zorin<sup>2</sup></a>
</div>
<div class="w-col w-col-3 w-col-small-2 w-col-tiny-6">
<a class="authors" href="http://web.eng.tau.ac.il/~raja/" target="_blank">Raja Giryes<sup>1</sup></a>
</div>
</div>
<div class="p2m_authors_list w-row center">
<div class="w-col w-col-3 w-col-small-2 w-col-tiny-6" style="margin-left: 25%;">
<a class="authors" href="https://cims.nyu.edu/gcl/daniele.html" target="_blank">Daniele Panozzo<sup>2</sup></a>
</div>
<div class="w-col w-col-3 w-col-small-4 w-col-tiny-6">
<a class="authors" href="https://www.cs.tau.ac.il/~dcor/" target="_blank">Daniel Cohen-Or<sup>1</sup></a>
</div>
</div>
<br>
<div class="div-block-10">
<div class="equal_v2">Tel Aviv University<sup>1</sup></div>
 
<div class="equal_v2">New York University<sup>2</sup></div>
</div>
<div>
<br>
<span class="center"><img width=50% src="images/output.gif"></span>
<!-- <span class="center"><img src="https://gist.github.com/ranahanocka/9513a4444d38ea419a3d6922f3563831/raw/4d0b90de8beff343de236e198726e24eeda15b68/lizard.gif" alt="lizard.gif"></span>-->
</div>
<!--start links -->
<div class="p2m_authors_list_single w-row">
<div class="w-col w-col-4 w-col-small-3 w-col-tiny-4">
<a class="authors" href="" target="_blank">
<a href="https://arxiv.org/abs/2105.01604" target="_blank"><i
class="far fa-4x fa-file text-primary mb-3 "></i></a>
</a></div>
<div class="w-col w-col-4 w-col-small-3 w-col-tiny-4">
<a class="authors" href="" target="_blank">
<a href="https://github.com/galmetzer/dipole-normal-prop" target="_blank"><i
class="fab fa-4x fa-github text-primary mb-3 "></i></a>
</a></div>
<div class="w-col w-col-4 w-col-small-3 w-col-tiny-4">
<a class="authors" href="" target="_blank">
<a href="soon.html" target="_blank"><i
class="fas fa-thumbtack fa-4x text-primary mb-3 "></i></a>
</a></div>
</div>
<div class="div-block-4 w-row">
<div class="w-col w-col-4 w-col-small-3 w-col-tiny-4">
<div class="text-block-2"><strong style="color:#18446c" class="icon-bold-text">Paper</strong></div>
</div>
<div class="w-col w-col-4 w-col-small-3 w-col-tiny-4">
<div class="text-block-2">
<strong style="color:#18446c" class="icon-bold-text">Code</strong>
</div>
</div>
<div class="w-col w-col-4 w-col-small-3 w-col-tiny-4">
<div class="text-block-2"><strong style="color:#18446c" class="icon-bold-text">Slides</strong></div>
</div>
</div>
</div>
<!-- end links -->
</div>
</div>
<div class="white_section">
<div class="w-container"><h2 class="grey-heading">Abstract</h2>
<p class="paragraph-3 the_text">
Establishing a consistent normal orientation for point clouds is a notoriously difficult problem in geometry processing,
requiring attention to both local and global shape characteristics.
The normal direction of a point is a function of the local surface neighborhood;
yet, point clouds do not disclose the full underlying surface structure.
Even assuming known geodesic proximity, calculating a consistent normal orientation requires the global context.
In this work, we introduce a novel approach for establishing a globally consistent normal orientation for point clouds.
Our solution separates the local and global components into two different sub-problems.
In the local phase, we train a neural network to learn a coherent normal direction per patch (i.e., consistently oriented normals within a single patch).
In the global phase, we propagate the orientation across all coherent patches using a dipole propagation.
Our dipole propagation decides to orient each patch using the electric field defined by all previously orientated patches.
This gives rise to a global propagation that is stable, as well as being robust to nearby surfaces, holes, sharp features and noise.
</p>
</div>
</div>
<!-- <div class="section the_section" data-anchor="slide1">
<div class="w-container"><h2 class="grey-heading">Video</h2>
Comming Soon!
</div>
</div> -->
<div class="section the_section" data-anchor="slide1">
<div class="w-container"><h2 class="grey-heading">Video</h2>
<video width="80%" controls>
<source src="images/vid.mp4" type="video/mp4">
</video>
</div>
</div>
<div class="white_section">
<div class="w-container"><h2 class="grey-heading">Dipole Normal Propagation Overview</h2>
<p class="paragraph-3 the_text" >
Locally, defining what is inside or the outside of the surface ((i.e., which direction the *surface* normal points)), is not well defined:
</p>
<div><span class="center" id="uncertain" ><img width=70% src="images/001_local_uncertain.png"></span></div>
<p class="paragraph-3 the_text">
The orientation of the surface in the indicated square, can not be determined without global context of the rest of the shape.
</br>
To alleviate this problem, we divide the orientation problem into two steps: Local and Global.
</p>
<p class="paragraph-3 the_text">
We start by orienting each local patch <i>coherently</i>: meaning that all normals in the patch should point <i>either</i> inside or outside the shape.
Locally, patches can be very complicated and non-convex, and constructing a hand crafted technique can be especially challenging.
Instead, we learn the correct orientation using a neural network,
trained on ground truth point clouds sampled from oriented meshes.
</p>
<p class="paragraph-3 the_text">
Initially, the points are oriented according to a reference point (yellow) placed at the center mass of the patch.
</p>
<div><span class="center"><img width="70%" src="images/majority_curve.png"></span></div>
<p class="paragraph-3 the_text">
Then, an ensemble of networks decides which points ought to be corrected.
Our network is trained to classify the incoherent point normals (in red).
After we <i>flip</i> the incoherent point normals, we obtain a coherently oriented patch.
</p>
<div><span class="center"><img src="images/patch_overview.jpg"></span></div>
<p class="paragraph-3 the_text">
After a collection of coherent patches is constructed,
we use a robust and stable dipole propagation technique to align the local patches <i>Globally</i>.
</p>
<p class="paragraph-3 the_text">
Starting from the flattest patch, a dipole field from each point is iteratively constructed and used to determine the orientation for the rest of the patches.
At each step of the algorithm, the patch that has the strongest interaction with the field is selected,
and flipped (if needed) to align with the field according to the majority of its points.
The field is constructed globally, meaning that all previously oriented patches contribute to the total dipole field, giving global context at each step of the propagation.
In addition, the network-predicted confidence scores are used as the dipole magnitude of each point,
to prevent accumulative propagation error from uncertain points.
</p>
<div><span class="center"><img width=50% src="images/star.gif"></span></div>
<p class="paragraph-3 the_text">
This however does not requires a long time to compute, because a closed solution for the field at each point is available.
Also, because the dipole field arise from a solution to the Poisson equation which is linear,
the contribution of patches is simply summed up to compute the total field.
Allowing to only compute the delta contributions of each newly oriented patch in the iterative propagation process.
</p>
</div>
</div>
<div class="white_section">
<div class="w-container"><h2 class="grey-heading">Acknowledgements</h2>
<p class="paragraph-3 the_text">
We thank <a class= "link" href="http://shihaowu.net/" target="_blank">Shihao Wu</a> for his helpful suggestions,
and are grateful for rendering help from <a class= "link" href="https://bioblog3d.wordpress.com/" target="_blank">Tamir Lousky</a>.
</br>
This work is supported by the European research council (ERC-StG 757497 PI Giryes),
and the Israel Science Foundation (grants no. 2366/16 and 2492/20).
<br>
This work was supported in part through the NYU IT High Performance Computing resources,
services, and staff expertise. This work was partially supported by the NSF CAREER award 1652515,
the NSF grants IIS-1320635, DMS-1436591, DMS-1821334, OAC-1835712, OIA-1937043, CHS-1908767, CHS-1901091,
a gift from Adobe Research, a gift from nTopology, and a gift from Advanced Micro Devices, Inc.
</p>
</div>
</div>
</body></html>
