WEBVTT 00:00.080 --> 00:01.120 Welcome back. 00:01.160 --> 00:10.800 Whenever you change any object detection model, you get a different shape, a different input and output 00:10.800 --> 00:11.520 shapes. 00:11.760 --> 00:16.080 You should update the object detector class. 00:16.360 --> 00:19.520 So this is the most important class. 00:19.520 --> 00:22.040 This is the core of your project. 00:22.080 --> 00:31.440 This is the core of your application that you should update whenever you have a different model okay. 00:31.480 --> 00:41.200 Remember we need those input and output shapes Float32 one 640 640 and three. 00:41.520 --> 00:46.680 The output 32 Float32 one five 8400. 00:46.680 --> 00:54.680 For those, we're going to update those variables inside the companion object. 00:54.720 --> 01:07.680 The SSD mobile net expects 300 by 300 input, while the YOLO let me write YOLO shape expects 640 by 01:07.800 --> 01:09.320 640. 01:09.360 --> 01:14.520 Okay, so remove this and write 640. 01:14.560 --> 01:17.200 The maximum detection number is ten. 01:17.200 --> 01:20.080 Default or maximum result. 01:20.200 --> 01:21.480 Uh, default. 01:21.480 --> 01:24.880 Maximum result can make it like ten, for example. 01:25.080 --> 01:29.720 And you can add the other parameters. 01:29.720 --> 01:36.480 Private constant output size equals to 8400. 01:36.520 --> 01:44.480 Private constant val num number of features equals to 84. 01:44.640 --> 01:47.160 Confidence index equals to. 01:47.200 --> 01:53.320 For the class probability index, start equals to five. 01:53.360 --> 01:55.280 This is the corresponds to. 01:55.320 --> 02:00.480 This number of features correspond to this and this number. 02:00.480 --> 02:01.640 So it's five. 02:01.680 --> 02:03.840 We are detecting only one class. 02:04.000 --> 02:08.920 And the output of the data data per detection is five. 02:08.920 --> 02:10.720 So this is five. 02:11.000 --> 02:14.400 We set the number of features equals to five. 02:14.600 --> 02:18.090 The input size is correct 8400. 02:18.130 --> 02:25.450 And this constant there is no need for it because only we have only one class to detect. 02:25.490 --> 02:28.490 Okay, now let's create the output. 02:28.610 --> 02:33.490 So here we have the output classes output scores output detections. 02:33.490 --> 02:34.850 We're going to remove them. 02:34.970 --> 02:41.890 We're going to create only one output buffer matching that one five 8400. 02:41.930 --> 02:48.130 For that let me create val output equals to array. 02:48.490 --> 02:52.490 And here array of number of features. 02:52.690 --> 02:57.530 And set the float array of the output size. 02:57.530 --> 03:01.690 Here the number of features is five. 03:01.730 --> 03:03.570 It's an array of five. 03:03.930 --> 03:09.890 And it's a float array of output size which is 8400. 03:09.930 --> 03:12.770 Here we created an array of size one. 03:12.970 --> 03:17.170 Inside it we create an array of number of features which is five. 03:17.490 --> 03:27.290 Then we created a float array that it is float array which is float32 output size, which is 8400. 03:27.330 --> 03:27.810 Okay. 03:27.970 --> 03:31.890 So an array inside an array inside an array. 03:32.250 --> 03:36.530 This is the output buffer matching the model output. 03:36.530 --> 03:38.850 This is very very important. 03:38.930 --> 03:47.010 Now let's create the input buffer input buffer private void input buffer. 03:47.130 --> 03:54.690 Remember that input buffer is one 640 640 and three. 03:54.730 --> 04:06.130 For that you use tensor buffer dot create fixed size and inside it start with int array of one. 04:06.290 --> 04:08.890 Matching this image size. 04:08.930 --> 04:12.690 The variable that we created before 640. 04:12.810 --> 04:18.010 Image size 640 and three, which is the last one. 04:18.130 --> 04:24.690 Remember we need to return Float32 one 640 640 and three three. 04:24.730 --> 04:27.370 Is the color channels RGB? 04:27.570 --> 04:32.420 What is the data type uh of our of our input buffer. 04:32.420 --> 04:35.140 It's a float 32 okay. 04:35.180 --> 04:38.380 So data type dot float32 okay. 04:38.620 --> 04:42.740 Let's use the input buffer and output buffer. 04:42.860 --> 04:44.780 Scroll down here. 04:44.780 --> 04:47.420 We don't need all of those. 04:47.580 --> 04:50.460 Let me scroll down to analyze. 04:50.580 --> 04:59.580 We have the background scoop launch detections and the listener scroll down here to the detect prepare 04:59.580 --> 05:06.500 image for model image proxy interpreter dot run for multiple input outputs. 05:06.660 --> 05:18.940 We need to remove it because it's not our case here the interpreter dot run pass the input which is 05:19.060 --> 05:25.940 input buffer and pass the output which is output okay, not the output map. 05:25.940 --> 05:28.020 It's called the output. 05:28.020 --> 05:31.660 You see guys this is the objects that we created. 05:31.700 --> 05:33.980 The output and input. 05:34.100 --> 05:40.580 Scroll up the input buffer and the end, the input buffer and the output. 05:40.580 --> 05:44.500 You can name it as output buffer in order to distinguish it. 05:44.540 --> 05:48.140 Okay, so let me use this output buffer. 05:48.180 --> 05:54.060 Scroll down here inside the text and output buffer okay. 05:54.140 --> 05:56.700 Now let's parse the detections. 05:56.700 --> 06:02.140 You can return parse detection result that it's a list of detection object. 06:02.260 --> 06:05.460 So here parse the detections. 06:05.460 --> 06:07.220 Parse detections. 06:07.220 --> 06:15.940 Also we can create a new function to distinguish it from parse detection result because it's completely 06:15.940 --> 06:16.740 different. 06:16.740 --> 06:25.140 So I'll comment it and parse YOLO detections okay. 06:25.180 --> 06:28.300 This is the new function we need to create. 06:28.300 --> 06:33.100 So here private function partial detections. 06:33.140 --> 06:37.300 List of detection result detection object. 06:37.300 --> 06:40.380 And in the next video we'll continue with this function.