PointFlow/docs/index.html
Grendel Yang 372215d427 Update
2019-08-10 17:24:20 -07:00

225 lines
11 KiB
HTML
Executable file

<!doctype html>
<html lang="en">
<head>
<!-- Required meta tags -->
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1, shrink-to-fit=no">
<!-- Bootstrap CSS -->
<link rel="stylesheet" href="https://stackpath.bootstrapcdn.com/bootstrap/4.3.1/css/bootstrap.min.css"
integrity="sha384-ggOyR0iXCbMQv3Xipma34MD+dH/1fQ784/j6cY/iJTQUOhcWr7x9JvoRxT2MZw1T" crossorigin="anonymous">
<title>PointFlow</title>
<!-- Global site tag (gtag.js) - Google Analytics -->
<script async src="https://www.googletagmanager.com/gtag/js?id=UA-111714927-1"></script>
<script>
window.dataLayer = window.dataLayer || [];
function gtag(){dataLayer.push(arguments);}
gtag('js', new Date());
gtag('config', 'UA-111714927-1');
</script>
</head>
<body class="container" style="max-width:780px">
<!-- Title -->
<div>
<div class='row mt-5 mb-3'>
<div class='col text-center'>
<p class="h2 font-weight-normal">PointFlow: 3D Point Cloud Generation with Continuous Normalizing Flows</p>
</div>
</div>
<!-- authors -->
<div class='row text-center h5 font-weight-bold pl-4 pr-4 mb-4'>
<a class="col-md-3 col-xs-6" href="http://www.guandaoyang.com"><span>Guandao Yang*</span></a>
<a class="col-md-3 col-xs-6" href="http://www.cs.cornell.edu/~xhuang/"><span>Xun Huang*</span></a>
<a class="col-md-3 col-xs-6" href="http://www.cs.cornell.edu/~zekun/"><span>Zekun Hao</span></a>
<a class="col-md-3 col-xs-6" href="http://mingyuliu.net/"><span>Ming-Yu Liu</span></a>
<a class="col-md-6 col-xs-6" href="http://blogs.cornell.edu/techfaculty/serge-belongie/"><span>Serge Belongie</span></a>
<a class="col-md-6 col-xs-6" href="http://home.bharathh.info/"><span>Bharath Hariharan</span></a>
</div>
<!-- affiliations -->
<div class='row mt-1 mt-2' >
<div class='col text-center'>
<p class="h5 font-weight-light">
<a class="ml-4" href="https://www.cornell.edu/"><span>Cornell University</span></a>
<a class="mr-4 ml-4" href="https://tech.cornell.edu/"><span>Cornell Tech</span></a>
<a class="mr-4 ml-4" href="https://www.nvidia.com/"><span>NVIDIA</span></a>
</p>
</div>
</div>
<div class='row text-center h5 font-weight-light pl-4 pr-4 mb-4'>
<span class="col-md-12 col-xs-12" style="color:#007bff">*Equal contribution</span>
</div>
<div class='row justify-content-center' style="position: relative; width: 100%;height: 0;padding-bottom: 33%; margin:0">
<img
class="img-fluid rounded mx-auto d-block"
src="assets/teaser_sm.gif" alt="airplane"
style="position:absolute; width:100%; height:100%; top:0px; left:0px;"
srcset="assets/teaser.gif 1536w"
sizes="(min-width: 768px) 1536px, 100vw"
>
</div>
</div>
<!-- Paper section -->
<div>
<hr>
<div class='row'>
<div class='col-md-3 col-sm-3 col-xs-12 text-center col-sm-3'>
<div class="row mt-4">
<a href="https://arxiv.org/pdf/1906.12320.pdf" style="max-width:200px; margin-left:auto; margin-right:auto">
<img src="assets/paper.jpg" alt="paper-snapshot" class="img-thumbnail" width="80%" style="box-shadow: 10px 10px 5px grey;">
</a>
</div>
<div class="row mt-4">
<div class="col">
<a class="h5" href="https://arxiv.org/abs/1906.12320" style="margin-right:10px">
<span>[Arxiv]</span>
</a>
<a class="h5" href="https://github.com/stevenygd/PointFlow" style="margin-right:10px">
<span>[Code]</span>
</a>
<a class="h5" href="assets/Bibtex.txt" target="_blank">
<span>[Bibtex]</span>
</a>
</div>
</div>
</div>
<div class='col-md-9 col-sm-9 col-xs-12'>
<p class='h4 font-weight-bold '>Abstract</p>
<p style='line-height:1;'>
As 3D point clouds become the representation of choice
for multiple vision and graphics applications, the ability to
synthesize or reconstruct high-resolution, high-fidelity point
clouds becomes crucial. Despite the recent success of deep
learning models in discriminative tasks of point clouds,
generating point clouds remains challenging. This paper
proposes a principled probabilistic framework to generate
3D point clouds by modeling them as a distribution of
distributions. Specifically, we learn a two-level hierarchy
of distributions where the first level is the distribution of
shapes and the second level is the distribution of points
given a shape. This formulation allows us to both sample
shapes and sample an arbitrary number of points from
a shape. Our generative model, named PointFlow, learns
each level of the distribution with a continuous normalizing
flow. The invertibility of normalizing flows enables
computation of the likelihood during training and allows us
to train our model in the variational inference framework.
Empirically, we demonstrate that PointFlow achieves stateof-
the-art performance in point cloud generation. We additionally
show that our model can faithfully reconstruct
point clouds and learn useful representations in an unsupervised
manner.
</p>
</div>
</div>
</div>
<!-- Video -->
<div>
<hr>
<div class='row text-center'>
<div class='col'>
<p class='h2 mr-3'>Video</p>
</div>
</div>
<div class='row mt-3 text-center center-block' style=" margin-left:auto; margin-right:auto; max-width:560px">
<div class='col ml-1 mr-1' style="position: relative; width: 100%;height: 0;padding-bottom: 56%;">
<iframe
src="https://www.youtube.com/embed/jqBiv77xC0M"
frameborder="0"
allow="accelerometer; autoplay; encrypted-media; gyroscope; picture-in-picture" allowfullscreen
style="position: absolute;width: 100%;height: 100%; left: 0; top: 0;"
>
</iframe>
</div>
</div>
</div>
<!-- Architecture, explaination -->
<div>
<hr>
<div class='row text-center'>
<div class='col'>
<p class='h2'>Brief Introduction to the Method</p>
</div>
</div>
<div class='row mt-3'>
<div class='col-md-7 col-sm-7 col-xs-12 mt-3'>
<img src="assets/pcl_is_a_sample.gif" class="img-fluid" alt="architecture">
</div>
<div class='col-md-5 col-sm-5 col-xs-12 align-middle mt-5'>
<p class="text-break">
Each shape can be viewed as a distribution of 3D points.
In such distribution, points on the shape have higher probability and are more likely to be sampled.
A point cloud can be viewed as a set of points sampled from such distribution.
</p>
</div>
</div>
<div class='row mt-3'>
<div class='col-md-5 col-sm-5 col-xs-12 align-middle mt-5'>
<p class="text-break">
We use two continuous normalizing flows (CNFs) to model the distribution of shapes, while each shape is a distribution of 3D points.
The latent CNF transforms a vector sampled from the shape prior to a latent shape vector.
The point CNF transforms 3D points sampled from the point prior to a point cloud conditioned on a shape vector.
</p>
</div>
<div class='col-md-7 col-sm-7 col-xs-12 mt-3'>
<img src="assets/inference-architecture.gif" class="img-fluid" alt="architecture">
</div>
</div>
</div>
<!-- Results, transformation -->
<div>
<hr>
<div class='row text-center'>
<div class='col'>
<p class='h2'>Visulaization of the Flow</p>
</div>
</div>
<div class='row'>
<div class='col'>
<img src="assets/flows.gif" class="img-fluid rounded mx-auto d-block" alt="flow-airplane-4">
</div>
</div>
</div>
<!-- Ack -->
<div>
<hr>
<div class='row mb-5 text-center'>
<div class='col'>
<p class='h2'>Acknowledgements</p>
<p>This work was supported in part by a research gift from Magic Leap. Xun Huang was supported by NVIDIA fellowship.</p>
</div>
</div>
</div>
<!-- Optional JavaScript -->
<!-- jQuery first, then Popper.js, then Bootstrap JS -->
<script src="https://code.jquery.com/jquery-3.3.1.slim.min.js"
integrity="sha384-q8i/X+965DzO0rT7abK41JStQIAqVgRVzpbzo5smXKp4YfRvH+8abtTE1Pi6jizo" crossorigin="anonymous"></script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/popper.js/1.14.7/umd/popper.min.js"
integrity="sha384-UO2eT0CpHqdSJQ6hJty5KVphtPhzWj9WO1clHTMGa3JDZwrnQq4sF86dIHNDz0W1" crossorigin="anonymous"></script>
<script src="https://stackpath.bootstrapcdn.com/bootstrap/4.3.1/js/bootstrap.min.js"
integrity="sha384-JjSmVgyd0p3pXB1rRibZUAYoIIy6OrQ6VrjIEaFf/nJGzIxFDsf4x0xIM+B07jRM" crossorigin="anonymous"></script>
</body>
</html>