Video-language pre-training has attracted considerable attention recently for its promising performance on various downstream tasks. Most existing methods utilize the modality-specific or modality-joint representation architectures for the cross-modality pre-training. Different from previous methods, this paper presents a novel architecture named Memory-augmented Inter-Modality Bridge (MemBridge), which uses the learnable intermediate modality representations as the bridge for the interaction between videos and language. Specifically, in the transformer-based cross-modality encoder, we introduce the learnable bridge tokens as the interaction approach, which means the video and language tokens can only perceive information from bridge tokens and themselves. Moreover, a memory bank is proposed to store abundant modality interaction information for adaptively generating bridge tokens according to different cases, enhancing the capacity and robustness of the inter-modality bridge. Through pre-training, MemBridge explicitly models the representations for more sufficient inter-modality interaction. Comprehensive experiments show that our approach achieves competitive performance with previous methods on various downstream tasks including video-text retrieval, video captioning, and video question answering on multiple datasets, demonstrating the effectiveness of the proposed method. The code has been available at https://github.com/jahhaoyang/MemBridge.

Video-language pre-training (VLP) has attracted increasing attention for cross-modality understanding tasks. To enhance visual representations, recent works attempt to adopt transformer-based architectures as video encoders. These works usually focus on the visual representations of the sampled frames. Compared with frame representations, frame patches incorporate more fine-grained spatio-temporal information, which could lead to a better understanding of video contents. However, how to exploit the spatio-temporal information within frame patches for VLP has been less investigated. In this work, we propose a method to learn tube tokens to model the key spatio-temporal information from frame patches. To this end, multiple semantic centers are introduced to focus on the underlying patterns of frame patches. Based on each semantic center, the spatio-temporal information within frame patches is integrated into a unique tube token. Complementary to frame representations, tube tokens provide detailed clues of video contents. Furthermore, to better align the generated tube tokens and the contents of descriptions, a local alignment mechanism is introduced. The experiments based on a variety of downstream tasks demonstrate the effectiveness of the proposed method.