Exploiting the local temporal information for video captioning

Abstract

Typical video captioning methods are developed based on the encoder-decoder architecture. To better exploit the local temporal information, e.g., details about objects and their corresponding actions, we propose a reinforcement learning based method to predict the adaptive sliding window size sequentially for better event exploration. More specifically, we introduce the single Monte-Carlo sample to approximate the gradient of reward-based loss function. And the self-critical strategy is employed to estimate baseline reward to diminish the variance of gradients. Moreover, temporal attention is utilized to selectively focus on a subset of temporal frame representations while generating each word. In addition, to better initialize the decoder’s state, we utilize the motion features extracted by 3D CNNs with mean pooling to endow the decoder with the prior knowledge of the entire video. To evaluate the proposed method, experiments are performed on three public benchmark datasets: Microsoft Video Description Corpus (MSVD), MSR Video to Text challenge (MSR-VTT) and Charades. The experimental results demonstrate the effectiveness of our method by comparing with state-of-the-art methods.