Our inspiration was the frustration while trying to separate subjects from their backgrounds in images without Photoshop.
What it does
Our algorithm was intended to automatically detect and remove the background from images taken on a phone or computer. However, due to limitations on machine learning training time and compute power, the algorithm does not fully understand images yet and performs interesting yet aesthetically interesting results.
How we built it
This was built by first making a data collector using an Intel real sense depth camera and an image converter to convert jpgs to pngs. This device was then walked around the building taking hundreds of images.Then the data collected trained a CNN or Convolution Neural Network to use the RGB camera data to predict the distance of objects in the scene. This distance estimator is then combined with user input to decide how much of the background is removed.
Challenges we ran into
Varying size images were difficult to adapt to our machine learning algorithm which has a fixed input and output size.
Accomplishments that we're proud of
We are proud of designing and starting to train a Neural Network in one day.
What we learned
We learned how to interface dynamic input sizes to a fixed neural network input. We also learned how to interface with hardware to collect data and process images.
What's next for Intelligent Image Assistant
Retraining the Convolutional Neural Network to better estimate distance with a larger data-set.