Inspiration
Have you ever been bored of your Youtube recommendations? We were and we couldn't find a way to get quality video recommendations without leaving the YouTube sandbox. That's where our project comes in. Our aim is to let you find something similar to what you are watching across multiple video hosting services with the convenience of the click of a button.
What it does
Our project uses a classifier trained on thousands of videos artificial intelligence to learn the key content of videos and thus recommend videos related to what you want to watch. This means that the project is not hindered by inaccurate titles of videos and is based on the content of the video itself.
How we built it
We built a custom data set scrapping 50 video related diverse sub-reddits to use to train the model. Each sub-reddit had at least 1000 videos, thus making the data-set large and diverse. Most of our data set had videos shorter than 5 minutes.
Challenges we ran into
The challenges listed below provided a great learning opportunity for us:
- Data gathering, this was a tough one due to the sheer amount of time it took to collect the information.
- Due to the size and nature of the data set, it took many hours for a model to be trained and the data set itself required custom augmentation.
Accomplishments that we're proud of
- We managed to collect and store a robust custom data set.
- We managed to augment our inference functions to run on a CPU efficiently.
What's next for the project
- We plan to increase the amount of videos covered in our data set in order to cater for a larger variety of video content.
- We also plan on making a web extension.
Log in or sign up for Devpost to join the conversation.