Multimodal Latent Representation Learning for Video Moment Retrieval

The rise of artificial intelligence (AI) has revolutionized the processing and analysis of video sensor data, driving advancements in areas such as surveillance, autonomous driving, and personalized content recommendations. However, leveraging video data presents unique challenges, particularly in the time-intensive feature extraction process required for model training. This challenge is intensified in research environments lacking advanced hardware resources like GPUs. We propose a new method called the multimodal latent representation learning framework (MLRL) to address these limitations. MLRL enhances the performance of downstream tasks by conducting additional representation learning on pre-extracted features. By integrating and augmenting multimodal data, our method effectively predicts latent representations, leveraging pre-extracted features to reduce model training time and improve task performance. We validate the efficacy of MLRL on the video moment retrieval task using the QVHighlight dataset, benchmarking against the QD-DETR model. Our results demonstrate significant improvements, highlighting the potential of MLRL to streamline video data processing by leveraging pre-extracted features to bypass the time-consuming extraction process of raw sensor data and enhance model accuracy in various sensor-based applications.