1 00:00:00,330 --> 00:00:01,500 Hi and welcome back. 2 00:00:01,530 --> 00:00:07,680 So in this section, we'll be taking a look at intersection with a union, which is one of the ways 3 00:00:07,680 --> 00:00:10,980 we can assess how well our object detector is performing. 4 00:00:11,490 --> 00:00:13,380 So let's take a look at this example here. 5 00:00:13,890 --> 00:00:18,240 So imagine we have a data set and humans living dataset. 6 00:00:18,600 --> 00:00:24,960 So the ground truth for this car, for the bounding box that humans have labeled is this green box here. 7 00:00:25,560 --> 00:00:26,940 So that's a ground truth. 8 00:00:27,060 --> 00:00:32,100 So what if a contractor proposes this red box hill? 9 00:00:32,730 --> 00:00:35,160 Oh, that's not too bad, but isn't it? 10 00:00:35,670 --> 00:00:39,060 But how do we assess the goodness of this proposal? 11 00:00:39,750 --> 00:00:40,830 So we'll take a look at that. 12 00:00:41,430 --> 00:00:48,510 So as you can see here from a glance, it looks like the red box covers 90 percent of the ground should 13 00:00:48,570 --> 00:00:50,620 original box, which is great. 14 00:00:50,640 --> 00:00:51,960 It looks pretty good, doesn't it? 15 00:00:52,710 --> 00:00:55,140 Of it is this coverage concept. 16 00:00:55,530 --> 00:00:57,120 Is that a good metric to use? 17 00:00:57,720 --> 00:01:00,210 It's actually not, because take a look at this. 18 00:01:00,810 --> 00:01:08,100 This box also covers 90 percent of the original box, and by extension, you can just draw a box over 19 00:01:08,100 --> 00:01:13,920 the entire image and tend to always cover what the ground truth boxes because overlapping with everything. 20 00:01:14,700 --> 00:01:16,110 So basically. 21 00:01:16,290 --> 00:01:20,370 So we solve this by using the intersection per union metric. 22 00:01:20,820 --> 00:01:22,470 So let's take a closer look at that. 23 00:01:22,650 --> 00:01:28,930 So how much of the correct area is covered by our predicted bounding box to do that? 24 00:01:28,950 --> 00:01:30,450 Let's take a look at this formula here. 25 00:01:30,960 --> 00:01:33,090 So what is the size of a union? 26 00:01:33,090 --> 00:01:36,810 The size of the union is this region here, the shaded region here? 27 00:01:37,380 --> 00:01:39,000 And what's the size of the prediction? 28 00:01:39,180 --> 00:01:41,400 That's this entire red box here. 29 00:01:41,850 --> 00:01:49,950 So as you can see, the shaded region occupies what looks like roughly to be 50 percent of the predicted 30 00:01:49,950 --> 00:01:50,400 box. 31 00:01:51,030 --> 00:01:55,290 So that's borderline acceptable, but not great, to be honest. 32 00:01:55,890 --> 00:01:56,760 So what? 33 00:01:56,820 --> 00:01:59,790 Let's get to let's take a look at what you basically is. 34 00:02:00,210 --> 00:02:04,870 Are you considers the ratio of the overlap to the size of the proposed box. 35 00:02:04,890 --> 00:02:06,840 That's what we just did here. 36 00:02:07,200 --> 00:02:11,460 And generally, an IOU over point five is considered acceptable. 37 00:02:11,910 --> 00:02:14,100 However, like I said, it's good, not great. 38 00:02:14,100 --> 00:02:19,020 We would want this box to be a lot closer to the car in this case here. 39 00:02:20,280 --> 00:02:25,950 So the higher the IOU, the better the bungling box, because that would mean the union that means do 40 00:02:25,950 --> 00:02:33,330 overlap between the original ground truth and the proposed box is quite high compared to the size of 41 00:02:33,480 --> 00:02:34,150 the prediction. 42 00:02:34,170 --> 00:02:36,300 So that's why we get a high value value. 43 00:02:37,050 --> 00:02:40,290 So I would use a very simple concept. 44 00:02:40,920 --> 00:02:42,810 Cooper posted a clip across clearly. 45 00:02:43,530 --> 00:02:50,130 Next, we'll take a look at the mean average precision, which is another metric that we can use to 46 00:02:50,460 --> 00:02:54,240 evaluate optical factors, and it's actually the mean metric we do use. 47 00:02:54,660 --> 00:02:56,790 So I'll talk about this in the next section. 48 00:02:56,940 --> 00:02:57,360 Thank you.