ADM-201 dump PMP dumps pdf SSCP exam materials CBAP exam sample questions

视频的帧插补技术 – 译学馆
未登录,请登录后再发表信息
最新评论 (0)
播放视频

视频的帧插补技术

AI Learns Video Frame Interpolation | Two Minute Papers #197

亲爱的学者朋友们
Dear Fellow Scholars,
这里是凯罗尔•佐尔奈-费希尔的两分钟论文
this is Two Minute Papers with Károly Zsolnai-Fehér.
借助当今的图像技术
With today’s graphics technology,
我们可以享受许多十分流畅的 由自己运用
we can enjoy many really smooth videos that were created
每秒60帧的技术拍摄的视频
using 60 frames per second.
我们也很喜欢 希望你早就注意到
We love it too, and we hope that you noticed
我们最新的一百集或者更多的视频
that our last hundred or maybe even more episodes
都是60赫兹的
have been available in 60hz.
然而 我们得到的往往是
However, it oftentimes happens that
那些每秒20至
we’re given videos that have anything from 20 to
30帧的视频
30 frames per second.
这意味着 若我们把它们展示在每秒60帧的时轴上
This means that if we play them on a 60 fps timeline,
这些帧中的一半甚至更多
half or even more of these frames
将不会提供任何新的信息
will not provide any new information.
我们为得到漂亮的慢动作 缓放这些视频时
As we try to slow down the videos for some nice slow-motion action,
这个帧频会更糟
this ratio is even worse,
生成的视频输出极不稳定
creating an extremely choppy output video.
幸运的是 有能够推测
Fortunately, there are techniques that are able to guess
在这些中间画面发生什么的技术 并告诉我们
what happens in these intermediate frames and give them to us.
这就是我们所说的帧插补
This is what we call frame interpolation.
我们在这一领域做了些实验
We have had some previous experiments in this area
试图制作出一个神奇的
where we tried to create an amazing slow
泡泡融合的慢动作视频
motion version of a video with some bubbles merging.
进行帧插补的一个简单又标准的方法
A simple and standard way of doing frame interpolation
被称为帧融合
is called frame blending,
即对两个最接近的已知画面的简单均分
which is a simple averaging of the closest two known frames.
更多的先进技术都是基于光流的
The more advanced techniques are optical flow-based,
即一种决定两帧画面之间
which is a method to determine what motions
发生了什么动作的方法
happened between the two frames,
并且基于那些知识创建出新画面
and create new images based on that knowledge,
在多数情况下能够生成更高质量的结果
leading to higher quality results in most cases.
这项技术用了卷积神经网络
This technique uses a convolutional neural network
以完成类似的东西
to accomplish something similar,
但最终 它并没有给我们一副画面
but in the end, it doesn’t give us an image,
而是一系列的卷积核心程序
but a set of convolution kernels.
这是一种在先前
This is a transformation that is applied to the previous
和后续的画面中被用以制作中间画面的转换
and the next frame to produce the intermediate image.
它不是画面本身
It is not the image itself,
而是制作画面的处方 你愿意的话
but a recipe of how to produce it, if you will.
我们从早前卷积方法中获取了不少乐趣
We’ve had a ton of fun with convolutions earlier,
我们使用它
where we used them to create beautiful subsurface
实时为半透明材料创造美丽的深层散射效果
scattering effects for translucent materials in real time,
我们忠诚的学者们记得 曾在某时
and our more loyal Fellow Scholars remember that at some point,
我也拉出我的吉他来展示
I also pulled out my guitar and showed
运用基于卷积的回响技术时
what it would sound like inside a church
它在教堂里听起来会怎么样
using a convolution-based reverberation technique.
链接在视频简介中
The links are available in the video description,
一定要去看看
make sure to check them out!
既然我们这里有了神经网络
Since we have a neural network over here,
这个训练理所当然地发生于
it goes without saying that the training takes place
大量的前后相邻的像对中
on a large number of before-after image pairs,
因此这个网络能够制成
so that the network is able to produce
这些卷积核心程序
these convolution kernels.
当然 为了验证这个算法
Of course, to validate this algorithm,
我们也需要获得一个基本事实
we also need to have access to a ground truth
与之比较 我们可以通过保留一些关于部分
reference to compare against –
中间画面的信息来得到它 因此我们有了这个准确的图像
we can accomplish this by withholding some information
这些图像还需要这个算法在
about a few intermediate frames
因此我们有了这个准确的图像
so we have the true images
这些图像还需要这个算法在
which the algorithm would have to reproduce
不看图像的前提下再生成
without seeing it.
有点像在已知答案的情况下
Kind of like giving a test to a student
给学生一个测试
when we already know the answers.
此处你可以看到这样的对比
You can see such a comparison here.
而现在呢 让我们看看这些效果吧
And now, let’s have a look at these results!
正如你所见 它们极为流畅
As you can see, they are extremely smooth,
并且这个技术在这些图象中保持了
and the technique retains a lot of high-frequency
许多高频率的细节
details in these images.
视频看起来也呈现暂时地连贯
The videos also seem temporally coherent,
这意味着它全无恼人的闪烁效果
which means that it’s devoid of the annoying flickering effect
此处改进的效果通过某种方法体现出来了
where the reconstruction takes place in a way
随后各帧皆有不同
that’s be different in each subsequent frame.
任何闪烁都没发生
None of that happens here,
这项技术是一项巨大财富
which is an excellent property of this technique.
为了此项技术的Pathon数据编码可行
The python source code for this technique is available
且免费用于非商业领域
and is free for non-commercial uses.
我已在简介中放上了链接
I’ve put a link in the description,
如果你已尝试这项技术并自身有所收获
if you have given it a try and have some results of your own,
请务必把它上传到评论区
make sure to post them in the comments section
或我们社交网站的讨论区
or our subreddit discussion.
链接在简介中
The link is available in the description.
谢谢您的观看和大力支持
Thanks for watching and for your generous support,
下期再见
and I’ll see you next time!

发表评论

译制信息
视频概述

视频融合和慢动作镜头中往往会遇到因频率不同而导致的画面丢失或失真问题,而帧插补技术有效解决了这种烦恼,可免费用于非商业领域!

听录译者

收集自网络

翻译译者

GreenT

审核员

审核员 V

视频来源

https://www.youtube.com/watch?v=T_g6S3f0Z5I

相关推荐