This thesis investigates the problem of learning video representations for video understanding. Previous works have explored the use of data-driven deep learning approaches, which have been shown to be effective in learning useful video representations. However, obtaining large amounts of labeled data can be costly and time-consuming. We investigate self-supervised approach as for multimodal video data to overcome this challenge. Video data typically contains multiple modalities, such as visual, audio, transcribed speech, and textual captions, which can serve as pseudo-labels for representation learning without needing manual labeling. By utilizing these modalities, we can train deep representations over large-scale video data consisting of millions of video clips collected from the internet. We demonstrate the scalability benefits of multimodal self-supervision by achieving new state-of-the-art performance in various domains, including video action recognition, text-to-video retrieval, and text-to-video grounding.
We also examine the limitations of these approaches, which often rely on the association assumption involving multiple modalities of data used in self-supervision. For example, the text transcript is often assumed to be about the video content, and two segments of the same video share similar semantics. To overcome this problem, we propose new methods for learning video representations with more intelligent sampling strategies to capture samples that share high-level semantics or consistent concepts. The proposed methods include a clustering component to address false negative pairs in multimodal paired contrastive learning, a novel sampling strategy for finding visually groundable video-text pairs, an investigation of object tracking supervision for temporal association, and a new multimodal task for demonstrating the effectiveness of the proposed model. We aim to develop more robust and generalizable video representations for real-world applications, such as human-to-robot interaction and event extraction from large-scale news sources.
| Identifer | oai:union.ndltd.org:columbia.edu/oai:academiccommons.columbia.edu:10.7916/mx7q-bf97 |
| Date | January 2023 |
| Creators | Chen, Brian |
| Source Sets | Columbia University |
| Language | English |
| Detected Language | English |
| Type | Theses |
Page generated in 0.0031 seconds