Homework: Collect an Image Dataset
Background
When designing for machine learning, and specifically computer vision, it helps to learn to see the world as a machine. Now that you have a basic understanding of what it takes to train an image recognition model, it is time to go back into the field to collect a more targeted dataset that you can use to prototype with, and to evaluate the feasibility of you product idea.
Learning Objectives
After completing this exercise you will be able to collect a small-scale street-level image dataset, and label it for the purposes of training a model.
Instructions
- Go outside and with a camera or your smartphone take photos of the various classes your envisioned product’s model should be able to recognize
- Collect an ample amount of images for each label – ideally, as a group, you end up with around 25-50 images per label, but more may be better.
- Also collect an equal amount of images for the “null” class (one that does not show the things you are interested in)
- Make sure these images are as varied as possible, so that the classifier will be very accurate at recognising this label (low false negatives / high true positives)
- Download the images to your local computer
- Label the images in the training set by putting each in a named folder for that label (i.e. class)
- Upload the dataset to a cloud service for easy use during follow-up exercises
Product
After completing this exercise you will have a dataset which you can use to train a model in Google Teachable Machine that approximates the model your envisioned responsible urban AI system will use.
Follow-up
In the next week’s session, you will pool your dataset with those of your group members, and use it to train and test a model.