WEBVTT 00:00.080 --> 00:01.000 Welcome back. 00:01.120 --> 00:05.160 Now let's talk about the input and output shapes of our model. 00:05.200 --> 00:11.480 We have YOLO input Float32 one 640 640 and three. 00:11.520 --> 00:17.000 This defines what kind of image you need to feed into the model. 00:17.240 --> 00:19.520 Float32 is the data type. 00:19.760 --> 00:26.120 The pixel values should be 32 bit floating point numbers. 00:26.160 --> 00:32.880 This is usually means your image pixels need to be normalized. 00:32.960 --> 00:37.960 Example scaled from 0 to 255. 00:38.080 --> 00:42.400 We need to normalize them from 0 to 1. 00:42.440 --> 00:46.160 Now let's talk about the array or the tensor. 00:46.280 --> 00:50.120 The shape or dimensions of the expected data array. 00:50.160 --> 00:54.520 It's read as batch size height. 00:54.560 --> 00:56.320 Width channels. 00:56.440 --> 01:03.770 First of all, the batch size number one here means That this is the batch size. 01:04.050 --> 01:10.130 This is or this model is configured to process one image at a time. 01:10.290 --> 01:17.170 640 and 640 is the height and the width of the image in pixels. 01:17.330 --> 01:19.810 Three the channels. 01:19.850 --> 01:24.530 The three color channels red, green and blue RGB. 01:24.810 --> 01:36.010 In simple terms, you must preprocess your input image to be 640 by 640 pixel RGB image, normalized 01:36.010 --> 01:41.050 to Float32 values and submitted as a batch of one. 01:41.170 --> 01:44.810 Now let's talk about the output tensor. 01:45.010 --> 01:49.650 This is the raw, unprocessed output of the YOLO model. 01:49.890 --> 01:57.530 It contains all the potential directions and needs to be parsed to be useful. 01:57.530 --> 01:58.690 Float32. 01:58.730 --> 02:05.180 The data type of the output values one, five and 808,400. 02:05.220 --> 02:05.860 The shape. 02:06.100 --> 02:08.260 This is the most complex part. 02:08.300 --> 02:16.220 It's best understood as batch size data per detection and total possible detections. 02:16.380 --> 02:24.460 Let's start with the batch size corresponds to one image we fed in and one image at a time. 02:24.500 --> 02:27.540 Five data per detection. 02:27.780 --> 02:35.780 This is the crucial part for each of the 8000 808,400 potential directions. 02:36.060 --> 02:39.100 The model outputs five values. 02:39.340 --> 02:46.900 The composition of these five values depends on the number of classes your model was trained to detect. 02:46.940 --> 02:51.860 8400 the total possible detections. 02:52.180 --> 02:58.780 This is the number of the anchor boxes or potential object locations. 02:58.780 --> 03:08.020 The model checks YOLO divides the image into a grid, and the checks for objects at each cell at different 03:08.020 --> 03:08.780 scales. 03:08.780 --> 03:15.900 For a modern YOLO version eight model, this number is typically 80 times 80 plus, 40 times 40 plus 03:15.900 --> 03:19.820 20 times 20 equals to 8400. 03:19.860 --> 03:23.380 And this is the total possible detections. 03:23.380 --> 03:27.020 The most important part is number five. 03:27.060 --> 03:30.620 We have the most common structure of these five values. 03:30.780 --> 03:33.180 We have center x. 03:33.220 --> 03:34.620 We talked about it. 03:34.620 --> 03:38.020 And center y center y. 03:38.380 --> 03:40.580 This is the second value. 03:40.820 --> 03:49.340 The third value is the height and the fourth value is the object Objectness uh score. 03:49.580 --> 03:52.740 And the fifth value is the class score. 03:52.980 --> 03:59.780 We're going to work with those values and data per detection in the Android studio. 03:59.780 --> 04:00.820 So don't worry. 04:01.060 --> 04:09.110 The x and y center x and y, the coordinates of the center of the detected bounding box width and height. 04:09.230 --> 04:13.950 Um, the dimensions of the detected bounding box and the object. 04:13.990 --> 04:20.910 Objective object score the model's confidence that this box contains any object at all. 04:20.990 --> 04:22.150 That's all. 04:22.750 --> 04:27.310 And that's for values with only one value left. 04:27.310 --> 04:32.030 It must be the class probability for that single class. 04:32.030 --> 04:34.870 There is no need for multiple class scores. 04:34.910 --> 04:49.150 Okay, so, uh, in short, your model takes a 640 by 640 image and outputs 8400 candidate boxes for 04:49.150 --> 04:57.550 a single object class, which you must then filter and clean up to get the final useful detections. 04:57.590 --> 04:58.070 Okay. 04:58.310 --> 05:01.910 This is what we're going to work with in Android.