Document Classification:
Reflection on Our Cat-Only Pet Door Model
What kind of data did you include in your training dataset and why?
For our training dataset, we used three different classes:
Orange and White Cats – To help the model recognize a common cat color pattern.
Black and Grey Cats – To include darker-colored cats and avoid bias toward only light-colored felines.
Groups of Cats – To test if the model can still recognize multiple cats entering at once.
We chose these categories to ensure the model could identify different types of cats, making it more inclusive and effective for cat owners with various breeds.
What other kind of data could have been helpful but maybe you couldn’t get in the short-term/for free?
Some useful but harder-to-find data included:
Cats of other colors (e.g., brown, Siamese, hairless cats) that didn’t fit neatly into our categories.
Nighttime images or low-light scenarios to improve recognition in the dark.
Obscured or partially visible cats (e.g., hiding behind objects) to test if the model can still recognize them.
Non-cat animals (dogs, raccoons, etc.) to improve accuracy in blocking unwanted guests.
Finding large, diverse datasets for free was challenging, but we could expand our dataset in the future by scraping images from free databases or using YouTube video clips.
Did your model work well for what you wanted?
Our model successfully recognized different types of cats based on color, meaning it worked well for many cat owners. However, some limitations became clear:
The model might struggle with cats of colors we didn’t include, like brown or spotted cats.
Non-cat animals weren’t part of the dataset, meaning the pet door might not correctly reject dogs or raccoons.
The “Groups of Cats” category could be tricky since multiple cats together might form unfamiliar shapes.
In what instances might your model not work very well?
If a cat outside our trained color groups (like a Siamese or tortoiseshell) tries to enter, it might not be recognized.
If the lighting is poor (e.g., at night), the model might misidentify or fail to recognize a cat.
If an animal that looks somewhat like a cat (e.g., a small dog) tries to enter, the model might mistakenly allow it in.
A printed image of a cat or a cat face on a screen might fool the model if it wasn’t trained to ignore these cases.
Project Link
https://teachablemachine.withgoogle.com/models/A799Jc12v/