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

基于AI的3D姿态估计技术突破——几乎实时! – 译学馆
未登陆,请登陆后再发表信息
最新评论 (0)
播放视频

基于AI的3D姿态估计技术突破——几乎实时!

AI-Based 3D Pose Estimation: Almost Real Time!

亲爱的学者朋友们
Dear Fellow Scholars,
这里是Károly Zsolnai-Fehér带来的的“两分钟论文”
this is Two Minute Paperswith Károly Zsolnai-Fehér.
本集要介绍一篇有关姿态估计的优秀新论文
This episode is about a really nice new paper on pose estimation.
所谓的姿态估计是指
Pose estimation means
我们提供人类的图像或者视频
that we have an image or video of a human as an input,
机器将会输出你在这看到的骨骼架构
and the output, should be, this skeleton that you see here
它把这个人当前的姿势展现给我们
that shows us what the current position of this person is.
听上去还行 但它在现实当中如何应用呢?
Sounds alright, but what are the applications of this, really?
它应用的范围很广
Well, it has a huge swath of applications,
比如 你们中的很多人经常听到的
for instance, many of you often hear about
电脑游戏和动画电影里的动作捕捉
motion capture for video games and animation movies,
它还在医学中被用来找出病人姿态的异常
but it is also used in medical applications for finding abnormalities in a patient’s posture,
用于动物行迹追踪 手语阅读
animal tracking, understanding sign language,
以及自动驾驶汽车的行人识别
pedestrian detection for self-driving cars,
除此之外还有很多
and much, much more.
所以如果我们能实时进行姿态估计
So if we can do something like this in real time,
那将会对许多现实应用非常有益
that’s hugely beneficial for many many applications.
然而 这同样是个巨大的挑战
However, this is a very challenging task,
因为人类的姿态表现多种多样
because humans have a large variety of appearances
而图像可能来自任何一个角度
images come in all kinds of possible viewpoints,
因此 姿态估计算法还要能解决遮挡问题
and as a result, the algorithm has to deal with occlusions as well.
这尤其的困难 我们看一下这里
This is particularly hard, have a look here.
这两个例子中 我们看不见左手的手肘
In these two cases, we don’t see the left elbow,
因而必须通过观察身体其他部分推断出来
so it has to be inferred from seeing the remainder of the body.
右边是我们的参考解决方案
We have the reference solution on the right,
从这里可以看出 这个新方法明显比
and as you see here, this new method is significantly closer to it
之前的做法更加贴近实际
than any of the previous works.
这个成果相当了不起
Quite remarkable.
这篇论文的主要研究
The main idea in this paper is that
对二维和三维姿势都适用
it works out the poses both in 2D and 3D
并且包含一种
and contains neural network
在保证动作一致性的前提下
that can convert to both directions between these representations
进行维度转换的神经网络模型
while retaining the consistencies between them.
首先 这项技术得到一个初始猜测
First, the technique comes up with an initial guess,
然后通过这些姿态转化网络
and follows up by using these pose transformer networks
对初始猜测进行修正
to further refine thisinitial guess.
这让一切都变得不一样了
This makes all the difference.
它不仅仅能提供高质量的结果
And not does it lead to high-quality results,
还比以前的算法耗时更短
but it also takes way less time than previous algorithms —
我们能在大约51毫秒内得到一个预测的姿态
we can expect to obtain a predicted pose in about 51 milliseconds,
也就是几乎每秒20帧
which is almost 20 frames per second.
这几乎是实时的
This is close to real time,
对我们前面提到过的许多现实应用来说
and is more than enough for many of the applications
这已经完全足够了
we’ve talked about earlier.
在硬件水平快速发展升级的时代
In the age of rapidly improving hardware,
这样的结果在质量和性能上
these are already fantastic results
都已经十分优秀了
both in terms of quality and performance.
不只是硬件
And not only the hardware,
论文水平也在以惊人的速度提升
but the papers are also improving at a remarkable pace.
活在这个时代可真好啊
What a time to be alive.
这篇论文有很详尽的评估段落
The paper contains an exhaustive evaluation section.
它将多种高质量解决方案进行对比
It is measured against a variety of high-quality solutions.
我建议你到视频简介里看一下
I recommend that you have a look in the video description.
但愿没人会在我实验室里安装一个
I hope nobody is going to install a system
每次我一打瞌睡 就会发出哔哔声的系统
in my lab that starts beeping every time I slouch a little,
但我还是很期待能从其他应用中获益的
but I am really looking forward to benefitting from these other applications.
感谢大家的观看以及大力支持
Thanks for watching and for your generous support,
下次见啦!
and I’ll see you next time!

发表评论

译制信息
视频概述

介绍一篇关于3D人体姿态估计的前沿研究论文,该技术可得到几乎实时的结果,是业内重大突破

听录译者

收集自网络

翻译译者

审核员

审核员_GS

视频来源

https://www.youtube.com/watch?v=F84jaIR5Uxc

相关推荐