A New Keyframe Selection Method Using the Concrete Distribution

A new neural network architecture to extract key frames is proposed. We combine ideas from both Learning to Explain: An Information-Theoretic Perspective on Model Interpretation, which is a work on instance-wise feature selection, and Concrete Autoencoders for Differentiable Feature Selection and Reconstruction, which is our previous work on unsupervised feature selection. Experiments are conducted on the BosphorusSign dataset, classification and reconstruction errors of the selected key frames are measured.