Skip to main content

Person Detection

An image classification model takes a grayscale image as input and detects whether a person is present in the image.

Screenshot 2022-03-31 at 4 29 00 PM

The first step of creating an ML pipeline is finding (or training) a Machine Learning Model that matches your application. Here, we have decided to choose the Person Detection model. We will start by knowing the model input/output information. So, click on the model node present inside the studio with the name of Person Detection. It will show the Input/Output information. This information will be used to build the ML Pipeline

image8

Comparing input with the image format: [batch_size, height, width, channels], we can see the model will take a 96 x 96 grayscale image (because the channel is 1). The input image type is u8, i.e., all the values will lie between [0, 255]. The output is an array of size [1, 1, 1, 3]. The model will return a list with scores for 3 labels. We will have to find the top score and its associated labels. Now, we have understood how the data will flow, let’s start with creating the ML pipeline.

  1. Drag an Image from the Inputs: We will set the output type to u8 and dimensions to [1, 99, 99, 1] because, as we saw above, this is the format needed by our model. The IMAGE takes the following arguments:

    • width - the image's width in pixels
    • height - the image's height in pixels
    • pixel-format - the format used by the pixels. Possible values are:
      • @PixelFormat::Grayscale
      • @PixelFormat::RGB

  2. Drag the Person Detection model from the left side panel.

  3. We have got the output score for all three classes from our model. We want to find out which 1 classes with the highest score so that later we can print that class name as output. To get the index of class with the highest confidence value, we will use most_confident_indices proc-block (a proc block which, when given a list of confidences, will return the indices of the top N most confident values). So, drag the most_confident_indices proc-block from the left side pan.

    image18

    Set input dimension same as output dimensions of the model [1, 1, 1, 3]. We want the three most confident values in the output, so set it to 1. You will have to set the count the same as the output dimensions.

  4. Till now, we have got an index of the most confident index. Next, we would like to assign a label to this index. We will follow the above approach and connect the output of the most_confident_indices node to the input of our label node. Fill the Input values the same as the output dimension of most_confident_indices. We will get a string as output, so we will have to set the output type to utf8 and dimensions to 1. In Properties, we will upload the labels of our model. You can find the label file here

    image6

  5. Finally, connect the output of the label proc-block node to the input of the Output node.