WEBVTT 00:00.080 --> 00:07.640 Before we start creating the partial detection function, let me check the prepare image for model if 00:07.640 --> 00:10.600 we need to do any changes here. 00:10.600 --> 00:16.840 Preparing the exact input tensor of the model, not unit eight one 300 303. 00:16.880 --> 00:18.040 This is for SSD. 00:18.320 --> 00:20.800 We need to prepare it for. 00:20.960 --> 00:30.720 For YOLO it's float 32 one 640 I am I'm repeating this because it's very important. 00:30.720 --> 00:39.200 This is the main difference between the models between YOLO, SSD and MobileNet and other models. 00:39.200 --> 00:40.680 So this is the bitmap. 00:40.880 --> 00:46.840 This is the tensor image passing the tensor image bitmap and the data type. 00:46.840 --> 00:48.920 It's not float32. 00:49.080 --> 00:56.400 Let me use this is for mobile net SSD and this is for yolo. 00:56.440 --> 01:03.740 YOLO use Float32 data type instead of using data type unsigned integer eight. 01:03.780 --> 01:06.820 We're going to use float 32. 01:07.020 --> 01:15.780 Then the image processor builder dot add resize operation the image size image size bilinear. 01:15.820 --> 01:18.340 This is the resize method. 01:18.500 --> 01:25.100 We can add rotation, image rotation and normalize operation and build. 01:25.100 --> 01:28.860 For YOLO there is normalization. 01:28.860 --> 01:32.260 So here YOLO normalization. 01:32.260 --> 01:41.420 So here add normalize operation from zero f 2 to 155 okay. 01:41.460 --> 01:46.180 This is the standard deviation alt plus enter to import the class. 01:46.180 --> 01:47.020 And here we go. 01:47.220 --> 01:48.260 Scroll down. 01:48.260 --> 01:50.140 This is the parse detection result. 01:50.140 --> 01:52.980 This is from the SSD MobileNet. 01:53.020 --> 01:56.860 It's different from it in YOLO. 01:56.860 --> 02:05.760 So let me remove this and let me start with the partial detections here. 02:05.760 --> 02:08.720 We need to return a list of detection object. 02:08.720 --> 02:17.280 So let's start with val results equals to mutable list of detection object. 02:17.440 --> 02:20.760 We need to store our results in a mutable list. 02:20.800 --> 02:26.400 Now for loop I in zero until the output size. 02:26.560 --> 02:27.840 What we need to do. 02:28.000 --> 02:33.520 We need to loop through this and get the boxes. 02:33.520 --> 02:35.040 How to get the boxes. 02:35.040 --> 02:40.600 As I told you, we have five data per detection. 02:40.920 --> 02:54.920 We have the X center, we have Y center, we have width, we have height and we have the confidence. 02:54.920 --> 03:04.380 Please write them down because we need to get for every detection those five Parameters to do that here. 03:04.620 --> 03:09.020 Let me start with x center equals to output. 03:09.060 --> 03:10.780 Output at index. 03:10.860 --> 03:15.660 In order to get the first index we start by zero zero. 03:15.860 --> 03:19.860 At an index I the output buffer. 03:20.060 --> 03:27.140 Then similar to that what we've done here, the y center is the second element. 03:27.180 --> 03:34.940 The width is the third element, the height is the fourth element, and the confidence is the fifth 03:34.980 --> 03:37.500 element at index I. 03:37.540 --> 03:38.100 Okay. 03:38.260 --> 03:48.180 The confidence index we created this variable before the confidence index you can remove it or you can 03:48.220 --> 03:50.620 use it like here. 03:50.780 --> 03:53.420 So scroll down to here. 03:53.580 --> 03:59.260 And instead of using four you can specify the confidence index okay. 03:59.540 --> 04:03.470 But for making it simple this is the confidence index. 04:03.470 --> 04:07.110 This is the height index with y center and x center. 04:07.150 --> 04:14.430 Those are the five parameters, the five data that we get from each detection. 04:14.550 --> 04:17.230 Now let me make some checks. 04:17.230 --> 04:23.110 If confidence greater than parameters and the confidence threshold. 04:23.110 --> 04:28.870 Or if the confidence thresholds the score threshold. 04:28.870 --> 04:30.070 What we need to do. 04:30.270 --> 04:36.670 We need to convert normalized center to pixel coordinates on preview surface. 04:36.670 --> 04:48.310 For that start with val x equals to x center minus width over two f times result view size dot width. 04:48.390 --> 04:57.710 Also similar to that, the y minimum for the height, the x maximum, and the y maximum. 04:57.990 --> 05:10.050 What are x maximum, y maximum, x maximum and y maximum That x minimum is the left side of the box, 05:10.250 --> 05:20.250 the y minimum is the top, the x maximum is the right, and the y maximum is the bottom. 05:20.290 --> 05:23.650 Okay, we are converting them. 05:24.010 --> 05:28.970 Convert normalized center to pixel. 05:29.010 --> 05:32.250 We converted that normalized center. 05:32.250 --> 05:35.970 This is the center of the detected object to pixels. 05:36.010 --> 05:41.370 Now let's draw the rectangle for detecting for the detected object. 05:41.490 --> 05:45.770 Rectangle equals to rect f left. 05:46.050 --> 05:47.210 Start by left. 05:47.250 --> 05:56.250 Curse in or curse at least zero f top at least zero f right at most. 05:56.290 --> 06:04.550 The result view dot width dot to float and the bottom curse at most result view dot view dot to float. 06:04.550 --> 06:06.350 This is the second step. 06:06.550 --> 06:09.190 This is the first step and this is the second step. 06:09.190 --> 06:12.590 The first step is to convert normalize center to pixels. 06:12.590 --> 06:21.750 Because as I told you in YOLO we we use normalized coordinates and normalize the normalized things. 06:21.750 --> 06:31.710 And in the second step let's draw the bounding box and applying clamping boundaries constraints to ensure 06:31.710 --> 06:36.310 the rectangle stays within the visible screen bounds okay. 06:36.510 --> 06:46.990 So using those curves at least and curse at most those clamps the bounding box coordinates to ensure 06:46.990 --> 06:52.910 they don't extend beyond the view's boundary zero to width or height. 06:52.950 --> 06:58.270 It's a safety measure to prevent drawing outside the screen. 06:58.310 --> 07:01.110 Okay, so this is a safety measure.