Guess the Celebrities in TV Shows!!

The objective of the project is to develop a system that takes a TV show or movie as input, and automatically detects and identifies all the celebrities shown in it. The system is useful for various applications such as video retrieval, automatic cast listing, etc. The task is more challenging than other face recognition tasks, because (a) compared to static pictures, videos have more motion blur and often poorer image quality, (b) face recognitions in videos require matching a character in various character poses, lighting conditions, facial expressions, makeup and hairstyle etc., which is difficult in most cases. The project is composed by a training phase and a runtime phase. During the training phase, a classifier based on CNN is trained on 14000 manually labeled faces taken from TV series Friends. In the training data set, each face is assigned a label, which is the name of the character shown in the image. And given a new image, the trained classifier is able to predict the shown celebrity and its probability. We experimented with approx. 13000 faces images and using an off-the shelf CNN with linear SVM achieved an overall accuracy of 72%. We further tested our classifier on the video of Friends movie, which was made in 2014 and achieved an overall accuracy of 49%. Our experiments confirm that good classification results can be achieved just by using an off the shelf CNN with SVM and also that aging significantly reduces the performance of the classifier.