kristian1108 / yt8m Goto Github PK

Drawing inspiration from recent advancements in video labeling and recommender systems, this project aims to develop a multi-staged approach to large-scale facial recognition in videos. In pursuit of this, models were developed and tested to understand context from videos, which in turn can be used to narrow down the likely set of faces appearing therein. In doing so, this model architecture could be useful in improving the effectiveness and efficiency of facial recognition software. The data used to develop these models came from the YouTube-8M and CelebA datasets. In order to support the ability to iterate and experiment quickly, a wrapper class for the YouTube-8M dataset was developed and shared publicly for others to use in their own research. The results of this project indicate promise for the concepts being tested, but further development is required to bring it to full fruition.