How computers learn to recognize objects instantly | Joseph Redmon

翻译人员: chunhua zhang 校对人员: 易帆 余

Ten years ago,

10年前，

computer vision researchers thought that getting a computer

计算机视觉研究者认为 要让一台电脑

to tell the difference between a cat and a dog

去分辨出一只猫和狗的不同之处

would be almost impossible,

几乎是不可能的，

even with the significant advance in the state of artificial intelligence.

即便是在人工智能已经取得了 重大突破的情况下。

Now we can do it at a level greater than 99 percent accuracy.

现在我们已经可以做到 让它的正确率在99%以上。

This is called image classification --

这个方法叫做图像分类——

give it an image, put a label to that image --

给它一张图，再给这张图贴上标签——

and computers know thousands of other categories as well.

通过这种方式，电脑就可以知道 数千种的分类。

I'm a graduate student at the University of Washington,

我是华盛顿大学的一名研究生，

and I work on a project called Darknet,

我致力于一个名叫“暗网”的项目，

which is a neural network framework

这是一个用来训练和测试 计算机视觉模型的

for training and testing computer vision models.

神经网络结构。

So let's just see what Darknet thinks

让我们来看看暗网是如何看待

of this image that we have.

我们手上的这张图片。

When we run our classifier

当我们在这张图片上

on this image,

运行识别器时，

we see we don't just get a prediction of dog or cat,

我们注意到，它不仅能判断出 图片上是猫是狗，

we actually get specific breed predictions.

还能给出它是哪个品种的预测。

That's the level of granularity we have now.

这就是我们目前所达到的粒度级别。

And it's correct.

而且它的预测是正确的。

My dog is in fact a malamute.

我的狗的确是一只 阿拉斯加雪橇犬。

So we've made amazing strides in image classification,

很明显，我们在图像识别上 取得了惊人的进步，

but what happens when we run our classifier

但是如果我们对这样一张图片上

on an image that looks like this?

运行识别器，会如何呢？

Well ...

看一下。。。。。

We see that the classifier comes back with a pretty similar prediction.

我们看到识别器给出了一个 非常相似的预测。

And it's correct, there is a malamute in the image,

而且是正确的，图中是有一只 阿拉斯加雪橇犬，

but just given this label, we don't actually know that much

但只使用这一个标签， 我们并不能真正的了解

about what's going on in the image.

这张图片里的故事。

We need something more powerful.

我们需要更强大的检测器。

I work on a problem called object detection,

我正在研究一个叫做 目标检测的问题，

where we look at an image and try to find all of the objects,

也就是我们尝试 将一张图上的所有目标物都找出来，

put bounding boxes around them

然后将它们分别框起来，

and say what those objects are.

再加上标注。

So here's what happens when we run a detector on this image.

这就是我们对这张照片 运行检测器时所发生的。

Now, with this kind of result,

基于这样的结果，

we can do a lot more with our computer vision algorithms.

我们可以用计算机视觉算法 做更多的事情。

We see that it knows that there's a cat and a dog.

我们发现，它知道 这里有一只猫和一只狗。

It knows their relative locations,

它知道它们的相对位置，

their size.

它们的大小。

It may even know some extra information.

它可能甚至还知道一些 额外的信息。

There's a book sitting in the background.

例如背景里有一本书。

And if you want to build a system on top of computer vision,

如果你想建立一个 基于计算机视觉的系统，

say a self-driving vehicle or a robotic system,

比如说无人驾驶汽车 或者机器人系统，

this is the kind of information that you want.

那么这就是你想要得到的那类信息。

You want something so that you can interact with the physical world.

你要一个能与物质世界互动的系统。

Now, when I started working on object detection,

当我最开始开展目标检测项目时，

it took 20 seconds to process a single image.

它要花20秒去处理一张图片。

And to get a feel for why speed is so important in this domain,

为了感受一下为什么速度 在这个领域是如此重要，

here's an example of an object detector

举一个例子，这是一个2秒钟

that takes two seconds to process an image.

就能处理一张图片的检测器。

So this is 10 times faster

这个检测器的速度要比

than the 20-seconds-per-image detector,

处理每张图需要20秒的 检测器快10倍，

and you can see that by the time it makes predictions,

你还可以看到 在它做出预测的时候，

the entire state of the world has changed,

被检测的世界已经发生变化了，

and this wouldn't be very useful

这对于一个应用来说

for an application.

是没有多大用处的。

If we speed this up by another factor of 10,

如果我们将它的速度再提升10倍，

this is a detector running at five frames per second.

这个检测器每秒可处理5张画面。

This is a lot better,

这就好很多了，

but for example,

但是，举个例子

if there's any significant movement,

如果有任何重大的移动 （它就反应不过来了），

I wouldn't want a system like this driving my car.

我可不想让这样的一个系统 来驾驶我的汽车。

This is our detection system running in real time on my laptop.

这是在我电脑上运行的 实时检测系统。

So it smoothly tracks me as I move around the frame,

当我在移动时，它能顺利地追踪我，

and it's robust to a wide variety of changes in size,

而且它强大到能适应不同的大小、

pose,

姿势、

forward, backward.

向前、向后的改变。

This is great.

很了不起。

This is what we really need

如果我们想要建造一个

if we're going to build systems on top of computer vision.

基于计算机视觉的系统， 那么这就是我们真正需要的。

(Applause)

（掌声）

So in just a few years,

仅仅是几年的时间，

we've gone from 20 seconds per image

我们就从每张图20秒，

to 20 milliseconds per image, a thousand times faster.

提升到了每张图20毫秒， 速度提高了1000倍。

How did we get there?

我们是如何做到的呢？

Well, in the past, object detection systems

事实上在过去，目标检测系统

would take an image like this

会将这张图片

and split it into a bunch of regions

分成很多小区域，

and then run a classifier on each of these regions,

然后在每一块区域运行一下识别器，

and high scores for that classifier

在识别器中获得最高分数（的输出）

would be considered detections in the image.

就会被认为是这张图片的检测结果。

But this involved running a classifier thousands of times over an image,

这涉及到要在一张图片上 运行数千次识别器，

thousands of neural network evaluations to produce detection.

以及数千次的神经网络评估 才能获得检测结果。

Instead, we trained a single network to do all of detection for us.

而现在，我们训练了可以做出 所有检测的单一网络。

It produces all of the bounding boxes and class probabilities simultaneously.

它能同时生成边界盒和类别概率。

With our system, instead of looking at an image thousands of times

使用我们的系统， 不需要为了生成检测结果

to produce detection,

去重复上千数次地看同一张图片，

you only look once,

“只看一次”就行了，

and that's why we call it the YOLO method of object detection.

这也是为什么我们称之为 目标检测的“YOLO”法。

So with this speed, we're not just limited to images;

有了这个速度，我们就 不仅限于识别图像了，

we can process video in real time.

还可以实时处理视频。

And now, instead of just seeing that cat and dog,

现在，我们不仅看到了猫和狗，

we can see them move around and interact with each other.

还能看到它们走来走去，互相嘻戏。

This is a detector that we trained

这是一个我们在微软的 COCO数据库上，

on 80 different classes

用80种不同种类的物品

in Microsoft's COCO dataset.

训练过的检测器。

It has all sorts of things like spoon and fork, bowl,

包含了各种东西， 像勺子、叉子、碗

common objects like that.

等常见物品。

It has a variety of more exotic things:

还有各种奇特的东西：

animals, cars, zebras, giraffes.

动物、汽车、斑马、长颈鹿。

And now we're going to do something fun.

现在我们要做点儿有趣的事情。

We're just going to go out into the audience

我们的摄像头将要对准观众区，

and see what kind of things we can detect.

看看能检测出什么。

Does anyone want a stuffed animal?

谁想要一个毛绒动物玩具？

There are some teddy bears out there.

观众席里有了一些泰迪熊。

And we can turn down our threshold for detection a little bit,

我们把检测阀值调低一点，

so we can find more of you guys out in the audience.

这样就可以找出更多的观众。

Let's see if we can get these stop signs.

看下我们能不能找出这些停车标志。

We find some backpacks.

我们找到了一些背包。

Let's just zoom in a little bit.

再放大一点。

And this is great.

非常棒。

And all of the processing is happening in real time

所有这些都是在电脑上

on the laptop.

实时处理的。

And it's important to remember

请大家记住：

that this is a general purpose object detection system,

这是一个通用的目标检测系统，

so we can train this for any image domain.

因此我们可以将它训练 用于任何领域的图像识别。

The same code that we use

我们在无人驾驶汽车中

to find stop signs or pedestrians,

用来发现停车标志、行人

bicycles in a self-driving vehicle,

和自行车的代码，

can be used to find cancer cells

同样可以用于在组织活检中

in a tissue biopsy.

找出癌细胞。

And there are researchers around the globe already using this technology

全球已经有很多研究者 正在利用这一技术

for advances in things like medicine, robotics.

在医学、机器人学等方面取得了进展。

This morning, I read a paper

今天早上，我刚读到一篇文章，

where they were taking a census of animals in Nairobi National Park

人们在内罗毕国家公园 对动物数量进行普查，

with YOLO as part of this detection system.

使用了YOLO作为检测系统的一部分。

And that's because Darknet is open source

这是因为暗网是一个开源项目，

and in the public domain, free for anyone to use.

在公共领域，任何人都可以免费使用。

(Applause)

（掌声）

But we wanted to make detection even more accessible and usable,

但是我们想要让检测器 能被更多人使用、也更好用，

so through a combination of model optimization,

因此通过结合模型优化，

network binarization and approximation,

网络二值化和近似法，

we actually have object detection running on a phone.

我们实际上已经可以 在手机上进行目标检测了。

(Applause)

（掌声）

And I'm really excited because now we have a pretty powerful solution

我真的很激动， 因为我们在这个低级的

to this low-level computer vision problem,

计算机视觉问题上 有了一个强大的解决方案，

and anyone can take it and build something with it.

而且任何人都可以 使用它来做些什么。

So now the rest is up to all of you

所以接下来就看所有在座的各位

and people around the world with access to this software,

以及世界上所有 能够使用这个软件的人了，

and I can't wait to see what people will build with this technology.

而我已经等不及想要看看， 人们会用这一技术造出什么来了。

Thank you.

谢谢。

(Applause)

（掌声）