WEBVTT 00:00.160 --> 00:06.200 The last function we need to create in this file is parse classification output. 00:06.320 --> 00:12.720 So here create private function parse classification output. 00:12.760 --> 00:15.120 It takes one parameter. 00:15.240 --> 00:16.760 Output buffer. 00:16.920 --> 00:19.560 It's of type tensor buffer. 00:19.560 --> 00:24.440 It returns a list of detection result class. 00:24.440 --> 00:29.280 Here start with val results equals to mutable. 00:29.400 --> 00:32.720 List of detection result. 00:32.720 --> 00:39.640 Also val output array equals to output buffer dot float array. 00:39.680 --> 00:46.320 Now I want from you to focus with me because there is a very important note here. 00:46.520 --> 00:48.080 We should pay attention to it. 00:48.120 --> 00:55.680 We're going to process classification model outputs and convert them into detection results. 00:55.680 --> 01:04.320 If you remember we stored the returning result of this parse classification output into the variable 01:04.320 --> 01:08.200 results and pass it to the listener on results method. 01:08.320 --> 01:17.200 And if you pay attention, the list of detection result is the returning type of the parse classification 01:17.240 --> 01:18.240 output function. 01:18.280 --> 01:22.360 Okay, now let's start to start the processing. 01:22.400 --> 01:34.120 For I in output array dot indices, if I less than labels size what we need to do. 01:34.280 --> 01:43.160 We need to start with val confidence equals to output array at index I. 01:43.360 --> 01:51.560 Then if the confidence greater than 0.5 f we need to add it to the results. 01:51.600 --> 01:54.960 Let me explain what we are doing here. 01:55.000 --> 02:00.370 As you know we are iterating through each element in the output array. 02:00.730 --> 02:07.170 Each element represents the confidence probability for a specific class. 02:07.210 --> 02:08.690 Okay, it's very simple. 02:08.690 --> 02:10.490 We are iterating over this array. 02:10.570 --> 02:12.250 Bounds checking. 02:12.410 --> 02:18.890 We are checking here if the labels dot size is is is accessible. 02:18.890 --> 02:27.930 And we ensure that we don't access beyond the available labels preventing index out of bounds exception. 02:27.970 --> 02:31.370 It's normal to now everything is normal. 02:31.570 --> 02:33.170 The val confidence. 02:33.170 --> 02:41.050 The output array at index I extract the confidence score of class at index I. 02:41.410 --> 02:50.450 We need to get the confidence at index I only processes classifications with 50% confidence. 02:50.650 --> 02:55.330 This filters out weak or unreliable predictions. 02:55.330 --> 03:00.850 So if there is a detection with 40% like a hard disk. 03:00.890 --> 03:02.170 We don't need it. 03:02.210 --> 03:08.290 We get classifications only higher than 50%. 03:08.290 --> 03:10.770 So this is the mean of it. 03:10.810 --> 03:11.170 Okay. 03:11.210 --> 03:12.050 It's very simple. 03:12.050 --> 03:19.570 If you need 60% higher 70% detection confidence, you can increase this number. 03:19.570 --> 03:20.250 For now. 03:20.290 --> 03:21.930 Keep it 50%. 03:21.970 --> 03:22.730 It's good. 03:22.770 --> 03:23.250 Okay. 03:23.650 --> 03:26.490 Now for classification models. 03:26.490 --> 03:29.410 We don't have bounding boxes. 03:29.410 --> 03:34.290 We'll create a dummy bounding box that covers the whole image. 03:34.290 --> 03:37.650 So for that starting with results dot add. 03:37.690 --> 03:41.170 And here set a detection result. 03:41.170 --> 03:47.450 Set the labels at label index confidence confidence and rect at rect f. 03:47.570 --> 03:50.290 Here let me draw this rect f. 03:50.490 --> 04:02.450 The left bound is for example zero f the left, the top 0F1, f right, and one f, but okay. 04:02.490 --> 04:06.370 This is a normalized coordinates of this results. 04:06.410 --> 04:11.050 The label class name from labels at index I. 04:11.090 --> 04:16.570 Confidence probability score from 0 to 1.0. 04:16.610 --> 04:25.850 So probability zero less than probability or confidence and one point that bounding box dummy rectangle 04:26.730 --> 04:32.170 covering the entire image okay so here dummy rectangle. 04:32.170 --> 04:38.370 By the way if we run the application and if there is any errors, any logical or drawing errors, we 04:38.370 --> 04:40.810 can modify those later on. 04:40.810 --> 04:41.290 Okay. 04:41.530 --> 04:45.010 Dummy rectangle covering entire image. 04:45.010 --> 04:48.770 We can replace it later on with with smaller okay. 04:48.970 --> 05:00.170 If the model detects hard disk with 0.8 confidence, it will throw it if the model detects like AirPods 05:00.450 --> 05:02.410 with confidence 0.3. 05:02.450 --> 05:03.970 It will not draw anything. 05:04.010 --> 05:04.490 Okay. 05:04.690 --> 05:05.050 Okay. 05:05.090 --> 05:06.170 It's very simple. 05:06.290 --> 05:08.650 Now I want from you to focus with me. 05:08.810 --> 05:13.330 The results don't add if the model detects if the model. 05:13.330 --> 05:14.570 Let me show you. 05:14.570 --> 05:22.890 If the model detects a hard disk with 0.8% confidence. 05:23.010 --> 05:24.810 What we need to pass here. 05:24.850 --> 05:26.650 What we need to add to the results. 05:26.650 --> 05:29.410 We're going to add detection. 05:29.570 --> 05:33.530 Result label is hard disk. 05:33.770 --> 05:38.810 So if we go to the labels dot text here we get hard disk. 05:38.810 --> 05:42.210 So the model will copy this name. 05:42.210 --> 05:47.050 This label passed to the first parameter of the detection result. 05:47.290 --> 05:54.850 Then the confidence is 0.8 and the bounding box would be a rect f. 05:54.850 --> 05:57.130 For example 0.1. 05:57.330 --> 05:59.900 Here we are drawing 0.0. 06:00.020 --> 06:02.940 So let me add 0.1 like this. 06:02.940 --> 06:05.660 You can draw it like this example. 06:05.660 --> 06:07.540 It's very good okay. 06:07.780 --> 06:11.180 So this is the model detection. 06:11.380 --> 06:16.660 And the detection result object will be added to the results array. 06:16.860 --> 06:23.660 Now let me return this results by specifying here the returning result. 06:23.700 --> 06:27.660 Returning result is result list. 06:27.700 --> 06:28.180 Okay. 06:28.460 --> 06:31.900 We need to move it outside till here. 06:32.060 --> 06:34.020 Let me show you guys. 06:34.500 --> 06:34.900 Oops. 06:34.940 --> 06:37.620 Let me remove it after this. 06:37.660 --> 06:38.780 And here we go. 06:39.060 --> 06:41.100 We have an extra one. 06:41.100 --> 06:44.500 And this should be for the list okay. 06:44.540 --> 06:48.660 So let me move it up and here we go okay. 06:48.740 --> 06:54.140 This is our function pass classification output.