Video Summarization: How to Use Deep-Learned Features Without a Large-Scale Dataset

This paper proposes a framework incorporating deep-learned features with the conventional machine learning models within which the objective function is optimized by using quadratic programming or quasi-Newton methods instead of an end-to-end deep learning approach which uses variants of stochastic gradient descent algorithms. A temporal segmentation algorithm is first scrutinized by using a learning to rank scheme to detect the abrupt changes of frame appearances in a video sequence. Afterward, a peak-searching algorithm, statisticssensitive non-linear iterative peak-clipping (SNIP), is employed to acquire the local maxima of the filtered video sequence after rank pooling, where each of the local maxima corresponds to a key frame in the video. Simulations show that the new approach outperforms the main state-of-the-art works on four public video datasets.