I wanted to try using Teachable Machine to detect different letters of the ASL alphabet. I realized that I needed to add more input images with my hand in different places (e.g. in the top left/right corners, in front of my face, closer to the camera) for better output.
I trained a model (view here) to detect if someone is properly wearing a mask or not. It’s a huge pet peeve of mine when someone is wearing a mask, but their nose holes are out, defeating the purpose of the mask. The classes the model detects are mask on, mask partially off, and mask off.
I originally tried to make a model detect if I was wearing different kinds of glasses, but I think that was too nuanced.
I created a “rock, paper, scissors” model using this tool. While this was a lighthearted exercise, it revealed how training ML data requires human touch to be most successful. When my idea initially emerged, I instantly thought the images would look like stock photo representations of ‘rock, paper scissors’, where the silhouette of the signals are recognizable to a human eye. However, when training the model, I found myself using natural gestures of rock paper scissors, resulting in many unrecognizable gestures visually. This made sure that the model was more accurate when interacting with it on camera, which was perhaps not achievable if the model was just made of stock photo images of rock paper scissors.
The model was not given an “environment” class which just composes images it is not supposed to recognize as other classes. As such, it recognizes empty space as car keys even if there are no car keys present.
I am pretty impressed by how quickly this model is created with reasonable results. Reminds me of the quick training models that azure custom vision offers.
I trained the model with my phone displaying the two colors, and the result also worked for real life objects. I also tried to test with intermediary colors within blue and yellow but that didn’t work well (might be influenced by phone screen reflection).
I was just playing around with Teachable Machine, and I found it hard to come up with something cool that I could do with just one hand or no hands (because one hand has to be holding down the webcam button). My first idea was to do hand puppets but I found out most hand puppets require two hands to make the animal, so I decided to do something very simple.
I have 2 classes: one with my hair down, and one with my hair tied back with a ponytail. I kept it simple to train so I could see what it would categorize any gray areas. For example, in the video I tested out one with all my hair down, then my hair moved back without tying it up, using my hand to gather my hair into a ponytail, and then I also tested out how much hair I need to keep in the front to register as class 1. This was a fun and simple experience to do for like 10 minutes!
This was really fun and silly, I was going to put music on the video clip but decided against it since I did it without music. This is also just me making a fool of myself so :pensive:. Oh well. I trained the detector on me dancing the Macarena with the main three lines of the dance and the choreography. For the record, I have never danced it before and I am not particularly good at it either.