WEBVTT 00:00.080 --> 00:00.960 Welcome back. 00:01.000 --> 00:05.240 We created the prepare image for model function. 00:05.440 --> 00:14.640 Now let's create the parts detection results function here under the previous function create function. 00:14.640 --> 00:17.960 Private function pass detection results. 00:17.960 --> 00:22.120 The returning type is a list of detection object. 00:22.280 --> 00:26.640 Start with val detected objects. 00:26.640 --> 00:29.960 It's a mutable list of detection object. 00:30.120 --> 00:40.560 Then val detection count equals to output number of detections at index zero. 00:40.800 --> 00:55.560 Output num detections at index 0.2, end dot course at most and get the maximum detection number. 00:55.600 --> 01:00.620 Maximum detection number which is Print ln. 01:00.860 --> 01:03.100 Number of detections. 01:03.340 --> 01:04.540 Detection count. 01:04.620 --> 01:12.260 This code extracts and limits the row detection results from the model output. 01:12.300 --> 01:20.580 Read detection count gets the number of valid detections from output detections. 01:20.700 --> 01:26.340 Output detection number array at index zero and apply safely. 01:26.620 --> 01:33.540 Limit ensures we don't need to exceed maximum detection number. 01:33.820 --> 01:43.260 The array bounds protection if we go up, this is included inside the companion object and set it to 01:43.460 --> 01:44.340 ten. 01:44.340 --> 01:47.780 So the maximum detection number is ten. 01:47.820 --> 01:50.300 You can specify it directly like this. 01:50.300 --> 01:55.100 But I prefer using the not not the hard coded. 01:55.380 --> 01:58.450 I prefer using the Variables. 01:58.490 --> 02:06.610 The initial results creates empty mutable list to store processed detections. 02:06.850 --> 02:11.490 This is the processed detections and debugging here. 02:11.530 --> 02:20.130 The printing the number of count for monitoring typically removed in production. 02:20.290 --> 02:22.930 This is our first step. 02:23.170 --> 02:24.770 Now we need that. 02:24.770 --> 02:30.290 The key points the bounds safe uses course at most. 02:30.330 --> 02:37.250 This is a new function that we introduced to prevent index out of bounds errors. 02:37.530 --> 02:40.170 So let me write this node down. 02:40.410 --> 02:47.330 The course at most prevents index out of bounds errors okay. 02:47.530 --> 02:54.130 This is the first step in converting row model outputs into usable detections objects. 02:54.170 --> 02:54.650 Okay. 02:54.890 --> 03:03.990 So in this in this function we are converting row model to detection object. 03:04.030 --> 03:16.790 Now for every detection for every integer I in zero until detection count what we need to do. 03:17.030 --> 03:24.910 We need to start with val score equals to output score. 03:25.230 --> 03:28.550 Get index zero at index I. 03:28.590 --> 03:39.550 This is the two dimensional array starting from row zero and index which is the the column at index 03:39.550 --> 03:40.350 I okay. 03:40.510 --> 03:45.310 What we need to do we need to filter by confidence. 03:45.310 --> 03:52.630 This is threshold we have if score less than parameters dot score threshold. 03:52.830 --> 04:01.500 So if we go to the score threshold As we said before, it's 0.6, meaning that every detected object 04:01.500 --> 04:13.180 with a score lower than 60%, it will be neglected, but higher every object out and and higher that 04:13.180 --> 04:14.100 has the. 04:14.140 --> 04:19.420 Any object has a score greater than 60%. 04:19.540 --> 04:26.060 Go and display a box around it and set it inside our detected objects. 04:26.180 --> 04:26.700 Okay. 04:27.140 --> 04:28.780 So here continue. 04:28.820 --> 04:33.740 Go and continue and make another iteration. 04:33.740 --> 04:40.660 If not, create a val label index output classes dot to integer. 04:40.780 --> 04:47.460 Then if the label index it's not in labels. 04:47.460 --> 04:51.060 Dot indices is what we need to do. 04:51.340 --> 04:53.820 We need to continue also. 04:53.840 --> 04:58.680 So here also we need to print or without using the print. 04:58.720 --> 05:02.120 We can make a log here but it's safe. 05:02.160 --> 05:12.800 Okay then remove this label index val label equals to labels at label index. 05:12.840 --> 05:17.000 Bounding box equals to output locations. 05:17.160 --> 05:23.080 This extracts the class, label and bounding box for each detection. 05:23.080 --> 05:32.160 Getting the label looks up the human readable class name using label index from output classes. 05:32.160 --> 05:38.160 If the label index equals to zero, label might be person. 05:38.160 --> 05:38.840 If we. 05:39.240 --> 05:50.400 If we notice that the SSD here MobileNet and we have the Coco database, we we can understand it better 05:50.400 --> 05:54.990 for from the labels and get bounding box here. 05:55.030 --> 06:00.550 Retrieving coordinates from output locations. 06:00.790 --> 06:08.630 It's a 2D array for a, typically a y minimum x minimum, y maximum x maximum. 06:08.750 --> 06:11.270 As normalized values between 0 and 1. 06:11.270 --> 06:22.030 If we go up, as I told you in the previous video, we talked about the shape here y minimum y x maximum 06:22.230 --> 06:25.670 or x minimum, y maximum and x maximum. 06:25.910 --> 06:28.150 This is the output array. 06:28.470 --> 06:33.390 So if we scroll down here we have the bounding box okay. 06:33.550 --> 06:35.310 It's very simple. 06:35.350 --> 06:38.950 Now let's continue with the SSD mobile. 06:38.950 --> 06:46.630 Net model that returning coordinates in the form that we have and draw the rectangle.