A Dynamic Multi-Label Video Classification based on Recurrent Neural Networks

Abstract

Deep CNNs are hand-designed owing to the complexity of their construction and the computational requirements of their training. Ensembling method has shown a effectively increasing performance on deep CNNs, although it consists redundancy of data and high end computational resources are needed. The existing work is based on genetic algorithm to generate and optimize the weights of a deep CNNs to perform image classification. In the existing work, the learning process on large data is slow, partial datasets are randomly picked for the evaluation to dramatically speed it up. The proposed RNNs learns to predict features, edges, lines, object boxes in the video. The RNNs do perform a repetitive task for every element of a sequence, with the output being always depended on the previous computations, it is a powerful tool for addressing time dependent data, such as the video data.