Machine Learning has become an inevitable part of our lives. We try to use some of these techniques to solve real word challenges.
What it does
Our model learns from the series of input videos featuring events from a basketball game to classify occurrences of various events on the testing data.
How we built it
We leveraged the AWS Sagemaker platform to program our code using the keras deep learning framework for python. We proposed two model architectures - one using raw images with a pre-trained Xception model and the other using raw images and features extracted from them in a parallel fashion in conjunction with pre-trained Xception model. The accuracy, recall, precision scores were calculated for the evaluation data.
Challenges we ran into
Concurrent access and server time out issues which was similar to a real time production environment. :) Handling biased data and extracting meaningful features.
Accomplishments that we're proud of
A universal model architecture that works across multiple actions and classes. The features extracted from the images provided further insights into model training. Activation Maps helped understand the models capturing the event.
What we learned
A lot! Deep learning models are stubborn kids and they only are effective if we use them efficiently. As they usually say, garbage in is garbage out and deep models follow them very precisely. And the highlight about deep models is it s hard to understand why they succeed or fail exactly.
What's next for Detect-It
We will try to make this model generic for all 10 actions given and convert this into a multiclass classification problem wherein we detect the presence or absence of any set of events.