1 00:00:00,240 --> 00:00:07,230 Hi and welcome to the lecture on object detection in this section will basically be taking a high level 2 00:00:07,230 --> 00:00:09,930 overview of the principles of object detection. 3 00:00:10,470 --> 00:00:11,670 So let's get started. 4 00:00:11,790 --> 00:00:14,610 So firstly, what is object detection? 5 00:00:15,150 --> 00:00:21,600 Well, so far we've extensively dealt with image classification tests that is determining an image as 6 00:00:21,600 --> 00:00:23,820 a whole to tell you what the image contains. 7 00:00:24,210 --> 00:00:27,400 An example of that is basically the sofar dataset. 8 00:00:27,420 --> 00:00:32,430 You can see these are all different images of airplanes, automobiles, birds and so on. 9 00:00:32,970 --> 00:00:39,210 So an image classification test, you're just telling us what the images we're not seeing, what is 10 00:00:39,390 --> 00:00:42,720 what are the objects in the image specifically like? 11 00:00:42,720 --> 00:00:48,150 I know, I know we're saying it's a call, but we're not actually saying it's a car and a wall or a 12 00:00:48,150 --> 00:00:51,900 bird and the tugboat behind it or a cage and so on. 13 00:00:52,380 --> 00:00:56,550 So let's take a look a bit more closely at what object detection gives us. 14 00:00:57,450 --> 00:01:04,830 So what if we wanted to individually classify the objects in an image like this? 15 00:01:04,980 --> 00:01:10,560 You can see now we have calls with a bounding box around each car in the scene. 16 00:01:10,620 --> 00:01:13,560 It's a bit hard to see the cause, but it actually works fairly well. 17 00:01:14,160 --> 00:01:19,180 Then you have people being drawn over with this box and the bus as well. 18 00:01:19,200 --> 00:01:23,820 So this is an example of an object detection algorithm, and it works quite well. 19 00:01:23,850 --> 00:01:29,400 This is actually, I believe this is a field of vision tree are used here and it's working quite well. 20 00:01:30,070 --> 00:01:32,010 And then you can see in that an example of it here. 21 00:01:32,220 --> 00:01:36,240 It's identifying individual objects and drawing this building in boxes around them. 22 00:01:36,660 --> 00:01:39,450 You have a ball club, mouse bottle of wine glass. 23 00:01:39,990 --> 00:01:41,480 This is quite cool, isn't it? 24 00:01:41,490 --> 00:01:46,770 An object detection is one of the most useful algorithms in computer vision. 25 00:01:47,250 --> 00:01:50,220 It has such a wide range of applications. 26 00:01:51,090 --> 00:01:54,660 And here's the last example here, which just gets a cup in an image here. 27 00:01:56,220 --> 00:01:58,890 So you may be thinking this is hard and it is hard. 28 00:01:59,600 --> 00:02:00,690 Take a look at this scene here. 29 00:02:00,990 --> 00:02:03,150 There's so many different objects to identify. 30 00:02:03,690 --> 00:02:11,370 So as you might rightfully infer, this is a challenging problem because we need to put two things localization 31 00:02:12,150 --> 00:02:16,050 that is determining where the boundaries of this object lie. 32 00:02:16,530 --> 00:02:20,430 So we need to figure out a way to draw a bounding box around object. 33 00:02:20,680 --> 00:02:21,480 We have identified. 34 00:02:22,230 --> 00:02:25,080 And then secondly, we need to classify what our object is. 35 00:02:25,440 --> 00:02:28,050 So this one's a bulb, this one's a cup and a mouse. 36 00:02:28,470 --> 00:02:32,490 So we need an algorithm, a computer vision, deep learning algorithm, preferably. 37 00:02:33,090 --> 00:02:36,040 Actually, that's pretty much the only way we can do this. 38 00:02:36,260 --> 00:02:42,900 We can use horror cascade classifiers, which you would have seen in the OpenCV U.S. to do detection. 39 00:02:43,410 --> 00:02:46,620 However, those don't work out too well in the real world. 40 00:02:47,010 --> 00:02:49,380 In many tests, and they're quite hard to trained. 41 00:02:50,280 --> 00:02:54,030 So let's take a look a deeper look at up to the object detectors. 42 00:02:55,140 --> 00:03:01,950 So optical directors are usually trained on one or more classes and is seek to produce a bounding box 43 00:03:01,950 --> 00:03:04,590 around the object class that's been identified. 44 00:03:05,250 --> 00:03:07,200 And there are many kinds of object detectors. 45 00:03:07,560 --> 00:03:12,150 There are some nonlinear deep learning based ones, as I mentioned, our cascade classifiers. 46 00:03:12,570 --> 00:03:15,930 And then also the sliding window at home, such histogram of gradients. 47 00:03:16,410 --> 00:03:22,200 However, these preceded the very cool, deep learning methods that we use nowadays. 48 00:03:22,740 --> 00:03:29,460 And these deep learning methods include things like RC and Estes YOLO, which is a very good detection 49 00:03:29,460 --> 00:03:30,660 and detection too. 50 00:03:30,690 --> 00:03:35,940 Well, it's a very good, efficient detect is actually quite good as well, and also rathlin it. 51 00:03:36,630 --> 00:03:40,830 And then in the deep learning world, we have two main types of public detectors. 52 00:03:41,250 --> 00:03:48,210 We have two shot, which uses two stages a region proposal stage and then a classification stage and 53 00:03:48,210 --> 00:03:53,010 then a single shot that was localization and image classification at once. 54 00:03:53,670 --> 00:03:58,770 So now let's take a look at the formal definition from Wikipedia for object detection. 55 00:03:59,340 --> 00:04:05,790 So object detection is a computer technology related to computer vision and image processing that deals 56 00:04:05,790 --> 00:04:13,260 with detecting instances of semantic objects of a certain class, such as humans buildings cause in 57 00:04:13,260 --> 00:04:14,820 digital images and videos. 58 00:04:15,660 --> 00:04:20,610 Well, research two means of public action, including face detection, which we have seen previously, 59 00:04:21,060 --> 00:04:28,080 and pedestrian detection nowadays loads of other applications you can use object detection for, and 60 00:04:28,080 --> 00:04:32,280 these include image retrieval, video surveillance and many, many others. 61 00:04:33,180 --> 00:04:39,090 So now that you've got a deeper understanding of object detection, I want to take a look at something 62 00:04:39,090 --> 00:04:40,920 called object segmentation. 63 00:04:41,340 --> 00:04:48,270 This is so that you understand the difference between object detection and segmentation tests, so it's 64 00:04:48,270 --> 00:04:49,470 related to object attention. 65 00:04:49,560 --> 00:04:53,970 However, it sometimes relies on some very different algorithms to do segmentation. 66 00:04:54,330 --> 00:04:56,700 So what exactly is object segmentation? 67 00:04:57,180 --> 00:04:59,340 Well, segmentation firstly involves. 68 00:04:59,780 --> 00:05:03,740 So level classification, and let's take a look at this example here. 69 00:05:04,280 --> 00:05:10,460 So what this means here is that this is a person here, and you can see every pixel of the person is 70 00:05:10,460 --> 00:05:11,840 highlighted in a red color. 71 00:05:12,290 --> 00:05:16,280 And so far, you can see that other people in the scene are also in this red color. 72 00:05:16,910 --> 00:05:20,420 Notice that the sidewalk has been identified as a class on its own. 73 00:05:20,840 --> 00:05:27,350 And it's all been shaded pink and the has been identified and also been chained like a pinkish docket. 74 00:05:27,380 --> 00:05:33,110 I think it's great and can see the cause being blue and trees being green. 75 00:05:33,890 --> 00:05:36,680 So that's what pixel level classification is. 76 00:05:37,070 --> 00:05:39,910 And this is an example of segmentation. 77 00:05:39,920 --> 00:05:42,080 So let's take a look at the differences here. 78 00:05:42,410 --> 00:05:49,940 So these objects in object recognition is image classification and object localization, which relates 79 00:05:49,940 --> 00:05:51,290 to object detection. 80 00:05:51,740 --> 00:05:55,040 And then afterwards, we have something called instant segmentation. 81 00:05:55,640 --> 00:06:01,550 Now it is two types of segmentation this one called semantic segmentation that we see on the right to 82 00:06:02,510 --> 00:06:06,710 notice the difference between semantic segmentation and instant segmentation. 83 00:06:07,250 --> 00:06:13,850 Now, an instant segmentation, we have the individual sheep being classified and shaded in different 84 00:06:13,850 --> 00:06:14,330 colors. 85 00:06:14,930 --> 00:06:17,030 They are all going to be the same class as sheep. 86 00:06:17,040 --> 00:06:21,110 But no, we have sheep one sheep, two sheep three and dog one. 87 00:06:21,770 --> 00:06:26,980 Whereas previously with semantic segmentation, which is what we saw in the previous image here, that 88 00:06:26,990 --> 00:06:27,590 was this one. 89 00:06:28,160 --> 00:06:35,210 This was an example of semantic segmentation loops, so you can see that all the classes are the same 90 00:06:35,210 --> 00:06:37,220 color and they all merge into one another. 91 00:06:37,910 --> 00:06:44,900 So if you want to do some sort of fine grained analysis and allow analysis of individual classes like 92 00:06:44,900 --> 00:06:50,660 this instance, segmentation that would be the method to use and you can see an example of it here. 93 00:06:51,080 --> 00:06:52,790 So this is what I wanted to talk about. 94 00:06:52,790 --> 00:06:58,160 So we have classification is just seeing that image has a cat classification. 95 00:06:58,160 --> 00:07:01,550 Plus localization is actually during a bounding box around a cat. 96 00:07:02,030 --> 00:07:06,560 Similarly, public detection is basically built on top of that same thing. 97 00:07:07,040 --> 00:07:12,110 However, object detection can no do multiple classes and do multiple boxes in an image. 98 00:07:12,560 --> 00:07:18,350 So generally, we always refer to this as object detection, and we don't really use classification 99 00:07:18,350 --> 00:07:21,060 and localization as an algorithm and computer vision. 100 00:07:21,470 --> 00:07:23,390 It's just the way to get to this. 101 00:07:24,020 --> 00:07:27,770 And then we have instant segmentation and also semantic segmentation. 102 00:07:28,100 --> 00:07:34,130 Instant segmentation is the more advanced vision of segmentation because we're now joined bounding boxes 103 00:07:34,130 --> 00:07:35,450 around individual cats. 104 00:07:36,710 --> 00:07:42,260 So like I said previously, object detection has a wide range of use cases. 105 00:07:42,890 --> 00:07:49,640 This is a screenshot I took from the rubble of sight removal as as a very cool computer vision company 106 00:07:49,640 --> 00:07:51,290 to have an annotation tool. 107 00:07:51,290 --> 00:07:56,480 They need models to provide more so many different models that you can train. 108 00:07:56,960 --> 00:07:58,010 It's really, really good. 109 00:07:58,010 --> 00:08:01,790 And the guys who are behind our workflow always making new updates. 110 00:08:01,790 --> 00:08:07,940 So I encourage you to check it out a little or a lot from these guys, and you can see how many different 111 00:08:07,940 --> 00:08:08,570 tests. 112 00:08:08,750 --> 00:08:15,740 Things like gas leak detection, flare stack monitoring, augmented reality, bean counting, which 113 00:08:15,740 --> 00:08:23,160 is just kind of niche to be fit traffic onto portable IDs, soccer player tracking self-driving cars 114 00:08:23,180 --> 00:08:29,870 the big one Remote Tech support, tennis line tracking, hard hat detection to something a lot of people 115 00:08:29,870 --> 00:08:30,770 are doing nowadays. 116 00:08:31,220 --> 00:08:38,510 Satellite imagery Logo ID Sushi Identifier I mean, that is an unlimited number of tests. 117 00:08:39,080 --> 00:08:42,500 It's like as if you're a computer vision startup. 118 00:08:42,500 --> 00:08:49,070 It's such a good thing just to go with an industry, talk to people, figure out what they would want 119 00:08:49,250 --> 00:08:55,940 to use cameras to identify or to detect, like so many different things out there that you wouldn't 120 00:08:55,940 --> 00:09:01,940 think about things like this and managing the equipment like just identifying like where suits and equipment 121 00:09:01,940 --> 00:09:04,220 is using cameras is very useful. 122 00:09:05,030 --> 00:09:08,390 So that's it for an intro into object detectors. 123 00:09:08,780 --> 00:09:12,110 Next, we'll take a look at the history of Early Object Texas. 124 00:09:12,470 --> 00:09:19,190 So we'll go to a quick overview on hard cascade classifiers and sliding window with histogram of gradients. 125 00:09:19,220 --> 00:09:21,350 So stay tuned for that lesson. 126 00:09:21,470 --> 00:09:22,250 Thank you for watching.