WEBVTT 00:00.080 --> 00:00.960 Welcome back. 00:01.000 --> 00:05.280 We understood our model, hierarchy and shapes. 00:05.560 --> 00:10.040 Now let's continue with the object detector class. 00:10.040 --> 00:19.520 And here under those variables start creating the SSD mobile net output structure. 00:19.560 --> 00:29.160 Start with private val output locations equals to array of size one. 00:29.400 --> 00:33.160 And what this array contains array. 00:33.480 --> 00:37.680 And it's of maximum detection number. 00:37.840 --> 00:42.920 And set this the float array of size four. 00:43.040 --> 00:46.880 Then private val output classes. 00:46.880 --> 00:57.280 It's an array one of float array maximum detection number private val output scores array of one float 00:57.320 --> 01:00.800 array and maximum detection number. 01:00.840 --> 01:03.240 Don't worry we're gonna clarify everything. 01:03.440 --> 01:08.290 Output detection number which is float array of size one. 01:08.330 --> 01:18.610 Actually, those are the variables that are used for handling the object detection model output. 01:18.730 --> 01:27.730 This is the classic output structure for an object detection model specifically designed to handle the 01:27.730 --> 01:31.410 float32 output we discussed. 01:31.450 --> 01:42.730 Okay, so let me write here the output structure output structure for an object which is float32 okay. 01:42.970 --> 01:51.970 The output locations bounding box coordinates in a normalized coordinates from 0 to 1. 01:52.170 --> 01:57.250 So the shape would be one maximum detection number for and. 01:57.570 --> 02:04.250 This purpose stores the bounding box coordinates for each detected object. 02:04.290 --> 02:15.370 Each box has four values, typically y minimum x minimum y maximum x maximum normalized to zero one 02:15.610 --> 02:20.410 or x center, y center width and height, which is normalized. 02:20.450 --> 02:20.930 Okay. 02:21.210 --> 02:25.810 So this is the bounding box in normalized coordinates. 02:26.010 --> 02:31.010 Here we have for example y maximum or y minimum. 02:31.010 --> 02:37.250 We start with minimum x minimum y max and x max. 02:37.370 --> 02:37.890 Okay. 02:38.050 --> 02:41.130 This is the first variable the output location. 02:41.130 --> 02:43.570 This is the for the box the bounding box. 02:43.890 --> 02:44.210 Okay. 02:44.250 --> 02:46.250 We need to display the box. 02:46.530 --> 02:49.650 So we have to create the boundaries. 02:49.810 --> 02:59.770 The boundaries in a in a for for uh for elements array or an array containing four elements. 02:59.810 --> 03:00.210 Okay. 03:00.410 --> 03:03.130 Now the output classes. 03:03.170 --> 03:10.410 Class IDs the shape would be one and maximum detection. 03:10.410 --> 03:18.330 Maximum detection number stores the class index for each detection. 03:18.530 --> 03:26.260 So this is a very important lesson because we need to understand the output also, and why we need them 03:26.260 --> 03:32.820 as a float32, and why we need to get those arrays and analyze them. 03:33.020 --> 03:40.100 So we get the first array that represents the y minimum x minimum y maximum x maximum. 03:40.100 --> 03:41.820 Those are for the bounding boxes. 03:42.060 --> 03:49.700 The second array is for storing the class index for each detection okay. 03:49.900 --> 03:55.220 So the integer class IDs stores as floats. 03:55.340 --> 04:02.060 For example 0.0, f 1.0, f 2.0 f, and so on. 04:02.180 --> 04:06.780 Okay, this is for the output classes. 04:06.780 --> 04:12.620 Now we have the output scores the confidence scores. 04:12.620 --> 04:16.460 So confidence scores shape one. 04:16.500 --> 04:27.590 Maximum detection number stores the confidence score for each detection floats between 0.0 and 1.0. 04:28.030 --> 04:32.830 So 0.6 refers for 60% confidence. 04:32.870 --> 04:37.110 This is the confidence that we used in drawing the boxes. 04:37.150 --> 04:40.190 If you remember here draw into canvas. 04:40.470 --> 04:47.310 This text is used to draw the object detection and get the score from it. 04:47.350 --> 04:58.270 Okay now we have the last one is the output number detections which is the valid detection count. 04:58.310 --> 05:09.670 For example, shape one tells how it tells you how many of the maximum detection number detections are 05:09.710 --> 05:11.470 usually valid. 05:11.670 --> 05:14.390 So how many valid detections. 05:14.390 --> 05:19.350 There are a lot of detections but they are not uh not sure. 05:19.630 --> 05:28.510 And if you remember here in the data, the detection parameters we set the threshold is 0.6. 05:28.670 --> 05:35.150 So if you uh, if you open your camera and direct it to the door. 05:35.350 --> 05:43.750 There is there is a possibility of detecting a window, but the window here is 0.4. 05:44.070 --> 05:48.870 It's detected, but it's that the score threshold is 0.6. 05:49.190 --> 05:55.390 So we have door 0.7, window 0.4. 05:55.710 --> 05:57.150 It's eliminated. 05:57.190 --> 05:59.710 We keep the door. 05:59.710 --> 06:03.590 So it will be 0.7 greater than 0.6. 06:03.590 --> 06:08.150 So the inference and the score is higher than 0.6. 06:08.190 --> 06:11.990 Go and display and draw the box around the object. 06:12.110 --> 06:15.550 This is the output detection only process. 06:15.550 --> 06:18.990 The first output number of detections. 06:18.990 --> 06:28.070 To end detections we need to get the highest rank, the highest inference, the highest score from the 06:28.070 --> 06:29.110 detections. 06:29.110 --> 06:37.190 So we learned about the output of an output structure for the SSD mobile. 06:37.230 --> 06:38.710 Net model.