Life is short. Too much goes in a busy day. In this world, where everyone is competing for our attention, it is hard to stand out in your walk of life. Whether you are a content creator, a videographer, a vlogger or just a regular everyday person; we want to maximize your visibility while minimizing the time spent doing so.
What it does
You give us videos and we'll give you a set of the best shots to choose from. ReelLife gives any content creator an easy way to find thumbnails, highlights, or even the perfect shot for your promotional material. By giving a curated list of top quality images we will cut down on time spent digging through footage looking for only the best shots.
How we built it
ReelLife is coded entirely in python, and runs on a Google Cloud virtual machine instance. Our Convolutional Neural Network was built using the Keras library with Tensorflow as the backend. We also used the cv2, pytube, sci-kit, and various other data science and python libraries. The dataset used to build the machine learning model was largely found on film-grab.com.
Challenges we ran into
Deciding what makes a shot "good" was not only a difficult problem for machines but also for humans. Our team had to all come together and try to build a criteria for what we were looking for in this dataset. Once we were all on the same page, we had to gather our data and go through the process of labeling it all. We wrote a number of simple scripts to aid in the efficiency of both gathering and labeling our data to ensure we could spend as much time as possible working on our project.
Accomplishments that we're proud of
ReelLife is able to take a whole 13 minute video of a fashion walk and return to us just the best images of the models. When we first put in a scene of Ocean's 11, we got back well lit frames of Brad Pit and George Clooney looking handsome on camera. At this point, it was clear that the hours spent meticulously sorting data had paid off to construct a generalized machine learning model.
What we learned
The biggest things we took away from this project had to do with: Data, Machine Learning, and Final output. When it came to data, a machine learning model is only as good as the data its been provided. It cannot learn and generalize a problem well if it does not have a diverse enough dataset.
What's next for ReelLife
We would like to expand to different categories and get a more robust data labeling system to allow for users to select for a greater variety of shots. Of course, with any machine learning application, it would be greatly beneficial to gather more data and spend more time finetuning model. As well as having a live beta site for users to submit their videos and have their highlights reeled in.