TED2017
Joseph Redmon: How computers learn to recognize objects instantly
喬瑟夫．瑞德蒙: 電腦是如何學習即時辨識物體的？
Filmed:
Readability: 4.5
2,471,805 views
10 年前，研究人員認為要讓電腦辨別出貓、狗，那根本是不可能的。今日，電腦視覺系統已經可以做到 99% 以上的辨識度。這是怎麼辦到的？喬瑟夫．瑞德蒙致力於YOLO（你只要看一眼）的開放原始碼物體辨識方法，它已經可以像閃電般的速度，辨識出圖片及影片中的物件——從斑馬到停止標誌。在這場令人驚嘆的演示中，瑞德蒙向我們展示了這項技術重要的進步里程碑，像是在自動駕駛車、機器人、甚至是癌症檢測上的應用。
Joseph Redmon - Computer scientist
Joseph Redmon works on the YOLO algorithm, which combines the simple face detection of your phone camera with a cloud-based AI -- in real time. Full bio
Joseph Redmon works on the YOLO algorithm, which combines the simple face detection of your phone camera with a cloud-based AI -- in real time. Full bio
Double-click the English transcript below to play the video.
00:12
Ten years ago,
0
825
1151
10 年前，
00:14
computer vision researchers
thought that getting a computer
thought that getting a computer
1
2000
2776
電腦視覺研究人員認為，
00:16
to tell the difference
between a cat and a dog
between a cat and a dog
2
4800
2696
要讓電腦辨別貓與狗的差別，
00:19
would be almost impossible,
3
7520
1976
幾乎是比登天還難，
00:21
even with the significant advance
in the state of artificial intelligence.
in the state of artificial intelligence.
4
9520
3696
即使用了相當先進的
人工智慧都很難辦到。
人工智慧都很難辦到。
00:25
Now we can do it at a level
greater than 99 percent accuracy.
greater than 99 percent accuracy.
5
13240
3560
現在我們可以把辨別的準確度
提升到 99% 以上。
提升到 99% 以上。
00:29
This is called image classification --
6
17680
1856
這技術叫做圖像分類——
00:31
give it an image,
put a label to that image --
put a label to that image --
7
19560
3096
給電腦看圖片，
並給圖片貼上標籤——
並給圖片貼上標籤——
00:34
and computers know
thousands of other categories as well.
thousands of other categories as well.
8
22680
3040
電腦還可以識別出
許多其它類別的東西。
許多其它類別的東西。
00:38
I'm a graduate student
at the University of Washington,
at the University of Washington,
9
26680
2896
我目前是華盛頓大學的研究生，
00:41
and I work on a project called Darknet,
10
29600
1896
我正在做一個專題叫做「暗黑網路」，
00:43
which is a neural network framework
11
31520
1696
它是一個用來訓練及測試
00:45
for training and testing
computer vision models.
computer vision models.
12
33240
2816
電腦視覺模型的神經網路架構。
00:48
So let's just see what Darknet thinks
13
36080
2976
所以，讓我們來瞧瞧暗黑網路
00:51
of this image that we have.
14
39080
1760
對我們照片識別能力的狀況。
00:54
When we run our classifier
15
42520
2336
當我們在這張照片上
00:56
on this image,
16
44880
1216
開啟我們的分類器，
00:58
we see we don't just get
a prediction of dog or cat,
a prediction of dog or cat,
17
46120
2456
可以看到電腦現在不只
在預測這是狗或貓，
在預測這是狗或貓，
01:00
we actually get
specific breed predictions.
specific breed predictions.
18
48600
2336
它實際上正在擷取特定品種的預測。
01:02
That's the level
of granularity we have now.
of granularity we have now.
19
50960
2176
這就是現在我們電腦的粒度等級。
01:05
And it's correct.
20
53160
1616
辨別正確。
01:06
My dog is in fact a malamute.
21
54800
1840
我的狗的確是隻雪橇犬。
01:09
So we've made amazing strides
in image classification,
in image classification,
22
57040
4336
所以，我們在圖像識別上
已經有了很大的進步，
已經有了很大的進步，
01:13
but what happens
when we run our classifier
when we run our classifier
23
61400
2000
但如果我們用識別器
01:15
on an image that looks like this?
24
63424
1960
來辨別這樣的照片呢？
01:19
Well ...
25
67080
1200
嗯……
01:24
We see that the classifier comes back
with a pretty similar prediction.
with a pretty similar prediction.
26
72640
3896
可以看到從分類器
得到的預測也相當類似。
得到的預測也相當類似。
01:28
And it's correct,
there is a malamute in the image,
there is a malamute in the image,
27
76560
3096
沒錯，圖片中有一隻雪橇狗，
01:31
but just given this label,
we don't actually know that much
we don't actually know that much
28
79680
3696
但它只給出一個標籤，
我們對這張照片的理解
還不是很完整。
還不是很完整。
01:35
about what's going on in the image.
29
83400
1667
01:37
We need something more powerful.
30
85091
1560
我們需要更強的東西。
01:39
I work on a problem
called object detection,
called object detection,
31
87240
2616
我正在研究一個問題，
叫做「物件偵測」，
叫做「物件偵測」，
01:41
where we look at an image
and try to find all of the objects,
and try to find all of the objects,
32
89880
2936
我們把一張照片中的
所有物體都找出來，
所有物體都找出來，
01:44
put bounding boxes around them
33
92840
1456
用邊界框把它們框起來，
01:46
and say what those objects are.
34
94320
1520
然後標示它們是那些東西。
01:48
So here's what happens
when we run a detector on this image.
when we run a detector on this image.
35
96400
3280
我們來看一下當我們在這一張圖片上
執行偵測軟體時，會發生甚麼事。
執行偵測軟體時，會發生甚麼事。
01:53
Now, with this kind of result,
36
101240
2256
現在，有了這類的結果，
01:55
we can do a lot more
with our computer vision algorithms.
with our computer vision algorithms.
37
103520
2696
我們就可以利用電腦視覺演算法，
幫我們做更多的事。
幫我們做更多的事。
01:58
We see that it knows
that there's a cat and a dog.
that there's a cat and a dog.
38
106240
2976
我們可以看到，
電腦知道圖片中有一隻貓和狗。
電腦知道圖片中有一隻貓和狗。
02:01
It knows their relative locations,
39
109240
2256
它知道牠們彼此的相對位置、
02:03
their size.
40
111520
1216
大小。
02:04
It may even know some extra information.
41
112760
1936
電腦甚至可能知道其它的資訊。
02:06
There's a book sitting in the background.
42
114720
1960
它也看到了背景中有一本書。
02:09
And if you want to build a system
on top of computer vision,
on top of computer vision,
43
117280
3256
如果你想要建立一個
基於電腦視覺系統的實用系統，
基於電腦視覺系統的實用系統，
02:12
say a self-driving vehicle
or a robotic system,
or a robotic system,
44
120560
3456
比如說，自動駕駛車或機械人系統，
02:16
this is the kind
of information that you want.
of information that you want.
45
124040
2456
這類就會是你想要的資訊。
02:18
You want something so that
you can interact with the physical world.
you can interact with the physical world.
46
126520
3239
你會想要一個可以
與實體世界互動的東西。
與實體世界互動的東西。
02:22
Now, when I started working
on object detection,
on object detection,
47
130759
2257
當我開始做物件偵測時，
02:25
it took 20 seconds
to process a single image.
to process a single image.
48
133040
3296
它要花 20 秒才能處理一張圖片。
02:28
And to get a feel for why
speed is so important in this domain,
speed is so important in this domain,
49
136360
3880
為了讓各位體會
為什麼這個領域這麼講究速度，
為什麼這個領域這麼講究速度，
02:33
here's an example of an object detector
50
141120
2536
我這邊做個執行物件偵測器的示範，
02:35
that takes two seconds
to process an image.
to process an image.
51
143680
2416
一張照片只要 2 秒的處理時間。
02:38
So this is 10 times faster
52
146120
2616
所以，比 20 秒一張的偵測器
02:40
than the 20-seconds-per-image detector,
53
148760
3536
快了 10 倍，
02:44
and you can see that by the time
it makes predictions,
it makes predictions,
54
152320
2656
各位可以看到，
在它識別圖像的過程中，
在它識別圖像的過程中，
02:47
the entire state of the world has changed,
55
155000
2040
周圍環境已經發生了變化，
02:49
and this wouldn't be very useful
56
157880
2416
但對一個應用軟體而言，
02:52
for an application.
57
160320
1416
這樣的速度是很鷄肋的。
02:53
If we speed this up
by another factor of 10,
by another factor of 10,
58
161760
2496
如果我們把另一個參數調升到 10 ，
02:56
this is a detector running
at five frames per second.
at five frames per second.
59
164280
2816
這個偵測器每秒
就可以識別 5 張圖片。
就可以識別 5 張圖片。
02:59
This is a lot better,
60
167120
1536
這樣好多了，
03:00
but for example,
61
168680
1976
但，假如，
03:02
if there's any significant movement,
62
170680
2296
移動很快的時候……
03:05
I wouldn't want a system
like this driving my car.
like this driving my car.
63
173000
2560
我可不想在我車上裝這樣慢的系統。
03:09
This is our detection system
running in real time on my laptop.
running in real time on my laptop.
64
177120
3240
這是在我筆電上運行的
即時偵測系統。
即時偵測系統。
03:13
So it smoothly tracks me
as I move around the frame,
as I move around the frame,
65
181000
3136
我在框框附近移動的時候，
它可以很順暢地追蹤著我，
它可以很順暢地追蹤著我，
03:16
and it's robust to a wide variety
of changes in size,
of changes in size,
66
184160
3720
而且，它可以根據不同的大小、
03:21
pose,
67
189440
1200
姿勢、
03:23
forward, backward.
68
191280
1856
前、後來做調整。
03:25
This is great.
69
193160
1216
太棒了。
03:26
This is what we really need
70
194400
1736
如果我們要建立一個
基於電腦視覺系統的實用系統，
基於電腦視覺系統的實用系統，
03:28
if we're going to build systems
on top of computer vision.
on top of computer vision.
71
196160
2896
這個才會是我真正想要的。
03:31
(Applause)
72
199080
4000
（掌聲）
03:36
So in just a few years,
73
204280
2176
所以，才幾年的時間，
03:38
we've gone from 20 seconds per image
74
206480
2656
我們從每 20 秒處理一張照片，
03:41
to 20 milliseconds per image,
a thousand times faster.
a thousand times faster.
75
209160
3536
進步到每張照片只要 20 毫秒，
快了 1000 倍。
快了 1000 倍。
03:44
How did we get there?
76
212720
1416
我們是如何辦到的？
03:46
Well, in the past,
object detection systems
object detection systems
77
214160
3016
過去，物件偵測系統，
03:49
would take an image like this
78
217200
1936
會把一張像這樣的照片，
03:51
and split it into a bunch of regions
79
219160
2456
分割成好幾個小區塊，
03:53
and then run a classifier
on each of these regions,
on each of these regions,
80
221640
3256
然後在每一個小區塊
運行分類器軟體，
運行分類器軟體，
03:56
and high scores for that classifier
81
224920
2536
相似度得分如果比較高
03:59
would be considered
detections in the image.
detections in the image.
82
227480
3136
會被識別器認為照片偵測成功。
04:02
But this involved running a classifier
thousands of times over an image,
thousands of times over an image,
83
230640
4056
但這樣一張圖片要執行
好幾千次的識別指令、
好幾千次的識別指令、
04:06
thousands of neural network evaluations
to produce detection.
to produce detection.
84
234720
2920
經過好幾千次的神經網路評估
才有辦法偵測出來。
才有辦法偵測出來。
04:11
Instead, we trained a single network
to do all of detection for us.
to do all of detection for us.
85
239240
4536
但我們不是這樣做，我們訓練了一個
網路模型來幫我們完成所有的偵測。
網路模型來幫我們完成所有的偵測。
04:15
It produces all of the bounding boxes
and class probabilities simultaneously.
and class probabilities simultaneously.
86
243800
4280
它可以同時產出邊界框
並同時對可能的結果進行評估。
並同時對可能的結果進行評估。
04:20
With our system, instead of looking
at an image thousands of times
at an image thousands of times
87
248680
3496
有了我們的系統，
你就不用一張圖片看了好幾千遍
你就不用一張圖片看了好幾千遍
04:24
to produce detection,
88
252200
1456
才能偵測出來。
04:25
you only look once,
89
253680
1256
你只要看一眼 (YOLO)，
04:26
and that's why we call it
the YOLO method of object detection.
the YOLO method of object detection.
90
254960
2920
所以我們簡稱這個
物件偵測技術為「YOLO」。
物件偵測技術為「YOLO」。
04:31
So with this speed,
we're not just limited to images;
we're not just limited to images;
91
259360
3976
所以，有了這樣的辨識速度，
我們不只可以偵測圖片；
我們不只可以偵測圖片；
04:35
we can process video in real time.
92
263360
2416
還可以處理即時的影片。
04:37
And now, instead of just seeing
that cat and dog,
that cat and dog,
93
265800
3096
現在各位看到的不是
貓、狗的靜態圖片，
貓、狗的靜態圖片，
04:40
we can see them move around
and interact with each other.
and interact with each other.
94
268920
2960
而是有牠們在移動、
互動的動態影片。
互動的動態影片。
04:46
This is a detector that we trained
95
274560
2056
這是我們用微軟 COCO 資料集裡
04:48
on 80 different classes
96
276640
4376
80 種不同的類別
04:53
in Microsoft's COCO dataset.
97
281040
3256
訓練出來的辨識器。
04:56
It has all sorts of things
like spoon and fork, bowl,
like spoon and fork, bowl,
98
284320
3336
它包含各種東西，
像是湯匙、叉子、碗
像是湯匙、叉子、碗
04:59
common objects like that.
99
287680
1800
這類的日常用品。
05:02
It has a variety of more exotic things:
100
290360
3096
它還有很多奇妙的東西：
05:05
animals, cars, zebras, giraffes.
101
293480
3256
動物、車子、斑馬、長頸鹿。
05:08
And now we're going to do something fun.
102
296760
1936
現在我們要進行一件好玩的事。
05:10
We're just going to go
out into the audience
out into the audience
103
298720
2096
我們會進到觀眾席，
05:12
and see what kind of things we can detect.
104
300840
2016
去看看能辨識到哪些東西。
05:14
Does anyone want a stuffed animal?
105
302880
1620
有誰要填充娃娃？
05:18
There are some teddy bears out there.
106
306000
1762
這邊還有一些泰迪熊。
05:22
And we can turn down
our threshold for detection a little bit,
our threshold for detection a little bit,
107
310040
4536
我們現在降低一下
對偵測結果的精確度的要求，
對偵測結果的精確度的要求，
05:26
so we can find more of you guys
out in the audience.
out in the audience.
108
314600
3400
這樣我們可以在觀眾席中
找到更多東西。
找到更多東西。
05:31
Let's see if we can get these stop signs.
109
319560
2336
我們來看看能不能偵測到停止標誌。
05:33
We find some backpacks.
110
321920
1880
我們有偵測到一些背包。
05:37
Let's just zoom in a little bit.
111
325880
1840
現在把鏡頭拉近一點。
05:42
And this is great.
112
330320
1256
這真的很厲害。
05:43
And all of the processing
is happening in real time
is happening in real time
113
331600
3176
所有的偵測流程
都可以在筆電裡即時呈現。
05:46
on the laptop.
114
334800
1200
05:49
And it's important to remember
115
337080
1456
更重要的是，
05:50
that this is a general purpose
object detection system,
object detection system,
116
338560
3216
這只是一個一般用的物件偵測系統，
05:53
so we can train this for any image domain.
117
341800
5000
我們還可以訓練它
辨別任何領域的照片。
辨別任何領域的照片。
06:00
The same code that we use
118
348320
2536
同樣的程式碼，
放在自動駕駛車裡，
放在自動駕駛車裡，
06:02
to find stop signs or pedestrians,
119
350880
2456
可以偵測到停止標誌、行人、
06:05
bicycles in a self-driving vehicle,
120
353360
1976
腳踏車，
06:07
can be used to find cancer cells
121
355360
2856
但放到組織切片
06:10
in a tissue biopsy.
122
358240
3016
就可以偵測出癌症細胞。
06:13
And there are researchers around the globe
already using this technology
already using this technology
123
361280
4040
現在全球有很多研究人員
已經開始在使用這項技術
已經開始在使用這項技術
06:18
for advances in things
like medicine, robotics.
like medicine, robotics.
124
366240
3416
做進一步的研究，
像是醫藥、機械人領域。
像是醫藥、機械人領域。
06:21
This morning, I read a paper
125
369680
1376
今天早上，我讀到一篇文章，
06:23
where they were taking a census
of animals in Nairobi National Park
of animals in Nairobi National Park
126
371080
4576
在奈洛比國家公園裡，
他們要對動物們進行統計調查，
他們要對動物們進行統計調查，
06:27
with YOLO as part
of this detection system.
of this detection system.
127
375680
3136
YOLO 就是其使用的
偵測系統的一部分。
偵測系統的一部分。
06:30
And that's because Darknet is open source
128
378840
3096
而這一切都是因為
暗黑網路是開放原始碼，
暗黑網路是開放原始碼，
06:33
and in the public domain,
free for anyone to use.
free for anyone to use.
129
381960
2520
在公眾領域，
任何人都可以免費使用。
任何人都可以免費使用。
06:37
(Applause)
130
385600
5696
（掌聲）
06:43
But we wanted to make detection
even more accessible and usable,
even more accessible and usable,
131
391320
4936
但我們希望偵測系統
可以更親民、更好用，
可以更親民、更好用，
06:48
so through a combination
of model optimization,
of model optimization,
132
396280
4056
所以在經過模型優化、
06:52
network binarization and approximation,
133
400360
2296
網路二值化及近似度化的整合後，
06:54
we actually have object detection
running on a phone.
running on a phone.
134
402680
3920
我們終於可以在手機上偵測物件。
07:04
(Applause)
135
412800
5320
（掌聲）
07:10
And I'm really excited because
now we have a pretty powerful solution
now we have a pretty powerful solution
136
418960
5056
而我真的相當興奮，因為我們現在
在低階的電腦影像處理問題上
有了相當強力的解決方式，
有了相當強力的解決方式，
07:16
to this low-level computer vision problem,
137
424040
2296
07:18
and anyone can take it
and build something with it.
and build something with it.
138
426360
3856
任何人都可以拿去並創造一些東西。
07:22
So now the rest is up to all of you
139
430240
3176
所以，接下來就看各位
07:25
and people around the world
with access to this software,
with access to this software,
140
433440
2936
以及全世界所有人
用這個軟體大展身手了，
用這個軟體大展身手了，
07:28
and I can't wait to see what people
will build with this technology.
will build with this technology.
141
436400
3656
我真的等不及想看看你們
用這項科技所做出來的產品。
用這項科技所做出來的產品。
07:32
Thank you.
142
440080
1216
謝謝。
07:33
(Applause)
143
441320
3440
（掌聲）
ABOUT THE SPEAKER
Joseph Redmon - Computer scientistJoseph Redmon works on the YOLO algorithm, which combines the simple face detection of your phone camera with a cloud-based AI -- in real time.
Why you should listen
Computer scientist Joseph Redmon is working on the YOLO (You Only Look Once) algorithm, which has a simple goal: to deliver image recognition and object detection at a speed that would seem science-fictional only a few years ago. The algorithm looks like the simple face detection of a camera app but with the level complexity of systems like Google's Deep Mind Cloud Vision, using Convolutional Deep Neural Networks to crunch object detection in realtime. It's the kind of technology that will be embedded on all smartphones in the next few years.
Redmon is also internet-famous for his resume.
Joseph Redmon | Speaker | TED.com