1 00:00:00,420 --> 00:00:06,540 And welcome back in this section, we'll take a look at non maximum suppression, which is a really 2 00:00:06,540 --> 00:00:10,680 good technique that can be used to clean up our bounding box proposals. 3 00:00:11,070 --> 00:00:12,060 So let's get started. 4 00:00:12,240 --> 00:00:20,730 So firstly, a very common issue that happens with all Americans at the doctors is an issue of overlapping 5 00:00:20,850 --> 00:00:22,230 bounding boxes like this. 6 00:00:22,560 --> 00:00:23,880 We have one car. 7 00:00:24,150 --> 00:00:30,810 Yet the object detector model predicts three of four boxes that overlap on the same object. 8 00:00:30,990 --> 00:00:36,060 That's a problem because if you were trying to do object counting, it's going to mess up your results, 9 00:00:36,420 --> 00:00:36,960 wouldn't it? 10 00:00:37,800 --> 00:00:44,310 So non depression is a really simple and effective technique to go from this to this. 11 00:00:44,550 --> 00:00:45,690 So let's see how it works. 12 00:00:46,140 --> 00:00:51,510 So firstly, you collect all the overlapping bounding boxes here, which are these three boxes here? 13 00:00:52,230 --> 00:00:55,980 Now we get the maximum probabilities associated with each box. 14 00:00:56,400 --> 00:00:58,640 That's a probability of the box belonging to a class. 15 00:00:58,660 --> 00:01:02,070 So imagine these are the probabilities here you can see from this diagram. 16 00:01:02,070 --> 00:01:03,900 So we have the initial box being here. 17 00:01:03,900 --> 00:01:07,530 Point it in the next box being point five and one point six. 18 00:01:07,980 --> 00:01:13,170 So we select the box with the highest confidence score, and this confidence score is basically the 19 00:01:13,170 --> 00:01:20,610 confidence that the classifiers and the object detector us making to see that this class is a call or 20 00:01:20,610 --> 00:01:22,170 whatever class we have in mind. 21 00:01:22,860 --> 00:01:28,590 So we add the one with the highest probability score here to something called the final proposal. 22 00:01:28,590 --> 00:01:30,030 This which is initially empty. 23 00:01:30,930 --> 00:01:36,570 And then we remove that box from the initial proposal list, which was the initial set of overlapping 24 00:01:36,570 --> 00:01:37,830 boxes that we see here. 25 00:01:38,700 --> 00:01:44,500 Then what we do, we calculate the IOU for the boxes in the initial proposal list, with the box in 26 00:01:44,500 --> 00:01:45,930 the final proposal list. 27 00:01:46,290 --> 00:01:54,240 So this green box here we calculate the IOU for 2.5 blocks over the green box and the point six box 28 00:01:54,240 --> 00:01:55,290 over the green box. 29 00:01:55,860 --> 00:01:58,170 Well, that gives us is a level of overlap. 30 00:01:58,290 --> 00:02:00,960 So just a way to measure how much those boxes overlap. 31 00:02:01,440 --> 00:02:07,890 So if you threshold, which is typically point five, but you can just as easily, if it's over that 32 00:02:07,890 --> 00:02:13,290 threshold, then we reject it because what's happening there in step five is that we're seeing these 33 00:02:13,290 --> 00:02:14,760 boxes overlap too much. 34 00:02:15,090 --> 00:02:20,970 So we should reject this and we basically do this for all the proposal, all the proposals in the initial 35 00:02:20,970 --> 00:02:21,510 list here. 36 00:02:21,960 --> 00:02:27,300 And that's how we end up with nice clean up objects back to a model upwards like this. 37 00:02:28,050 --> 00:02:31,500 So that's basically it for non non maximum suppression. 38 00:02:31,950 --> 00:02:37,950 It gets us NIMH non maximum suppression because all of the boxes that are not the maximum confidence 39 00:02:37,950 --> 00:02:39,690 score are basically suppressed. 40 00:02:39,840 --> 00:02:41,910 So it's basically it's self-explanatory name. 41 00:02:42,690 --> 00:02:44,060 So we'll stop there. 42 00:02:44,880 --> 00:02:51,540 And now I'll start giving you a tour of all the different object detector families starting first with 43 00:02:51,960 --> 00:02:56,580 our CNN's, which also includes Fast or CNN's at a faster RC and ends. 44 00:02:57,060 --> 00:02:58,850 So I'll see you in the next section. 45 00:02:58,860 --> 00:02:59,640 Thank you for watching.