A convolutional temporal encoder for video caption generation

Abstract

We propose a convolutional temporal encoding network for video sequence embedding and caption generation. The mainstream video captioning work is based on recurrent encoder of various forms (e.g. LSTMs and hierarchical encoders). In this work, a multi-layer convolutional neural network encoder is proposed. At the core of this encoder is a gated linear unit (GLU) that performs a linear convolutional transformation of input with a nonlinear gating, which has demonstrated superior performance in natural language modeling. Our model is built on top of this unit for video encoding and integrates several up-to-date tricks including batch normalization, skip connection and soft attention. Experiment on two large-scale benchmark datasets (MSAD and M-VAD) generates strong results and demonstrates the effectiveness of our model.