WEBVTT 00:00.120 --> 00:01.000 Welcome back. 00:01.040 --> 00:03.440 Let's continue with the detect function. 00:03.440 --> 00:12.120 And here under the image processor variable let me create another variable called tensor image. 00:12.120 --> 00:20.360 Here we have the tensor image Float32 tensor image load and tensor image equals to image processor dot 00:20.400 --> 00:21.240 process. 00:21.280 --> 00:31.280 The first line creates an image container, a TensorFlow Lite image container that holds image data 00:31.280 --> 00:35.080 in a format that model can can understand. 00:35.080 --> 00:44.560 So that tensor image is part of the TensorFlow Lite support library, and data type means each pixel 00:44.560 --> 00:49.480 will be stored as 32 bit floating point number. 00:49.720 --> 00:54.040 This matches most models input type. 00:54.080 --> 01:03.120 Essentially, it's an object wrapper that will hold your image data pixels in the model's expected tensor 01:03.220 --> 01:06.820 Format and the tensor image here. 01:07.020 --> 01:14.860 Tensor image dot load loads your Android's bitmap image into a tensor image. 01:14.860 --> 01:23.780 Converts the bitmaps pixel data into 32 float Float32 tensor format inside the object. 01:23.820 --> 01:31.300 Now the tensor image holds your row image but not yet preprocessed. 01:31.300 --> 01:38.780 Still has original size and pixel value 0 to 25 255. 01:38.820 --> 01:45.260 That tensor image equals to image pre-processor dot process. 01:45.300 --> 01:54.180 This passes the loaded image through your previously defined image processor pipeline. 01:54.180 --> 02:02.460 This will resize it to match your model's required input width and height, and normalizes the pixel 02:02.460 --> 02:09.160 values from zero 25 to 255 to 0 to one. 02:09.160 --> 02:18.040 Okay, so here I am, passing the image to the image processor to process and resize it and normalize 02:18.040 --> 02:22.600 it to fit our demands, our model demands. 02:22.800 --> 02:31.200 The result is a new processed image tensor image ready to be fed into your model. 02:31.240 --> 02:34.880 Okay, so this is a very important thing. 02:35.040 --> 02:42.680 We need to process the image, convert it into a tensor image to pass it and fed to your model. 02:42.680 --> 02:49.600 Then let's continue with creating the input buffer by creating val input buffer. 02:49.600 --> 02:52.840 And we talked about the input buffer in the previous videos. 02:53.040 --> 02:55.240 Tensor image dot buffer. 02:55.400 --> 02:58.040 Creating the output buffer. 02:58.080 --> 03:04.040 Let's create the output buffer based on your model's output size and shape. 03:04.160 --> 03:11.310 So val output buffer equals to tensor flow buffer. 03:11.430 --> 03:16.110 Tensor buffer dot create fixed size. 03:16.430 --> 03:20.270 We need to pass the parameters to parameters. 03:20.390 --> 03:23.150 The first parameter is the shape. 03:23.150 --> 03:33.030 It's an int array of one two and the data type float32 the output buffer. 03:33.070 --> 03:40.710 Think about it as a container, an output container to store the model's prediction result. 03:40.830 --> 03:49.030 That tensor buffer a class that holds a numeric tensor data used for input outputs in TensorFlow Lite 03:49.030 --> 03:59.990 and that create fixed size, is a function defining the shape of the output tensor one two, which which 03:59.990 --> 04:08.310 means that one sample with two output values, for example x and y coordinates for two prediction scores. 04:08.550 --> 04:17.450 So here we are creating a Function, a function that returns the input array of one and two. 04:17.490 --> 04:25.210 Defining the shape of the output tensor one sample with two output values. 04:25.250 --> 04:32.970 Okay, so I want from you to understand what those arrays and numbers and elements mean. 04:32.970 --> 04:42.490 One sample with two output values x and y are coordinates of two prediction scores and data type Float32 04:42.530 --> 04:43.650 means the output. 04:43.690 --> 04:53.250 The outputs are floating point numbers, so this line creates an empty result container with shape one, 04:53.450 --> 05:00.370 two and float values ready to receive the model's output after your call. 05:00.850 --> 05:08.010 This is how we define the output buffer and how to process the image.