Contextual Hashing with Decoupled Localization for Efficient Video Corpus Moment Retrieval

The aim of Video Corpus Moment Retrieval (VCMR) is to locate temporal moments within a large video corpus on the basis of textual queries; however, it faces challenges in balancing retrieval efficiency and localization accuracy. Existing approaches typically rely on high-dimensional continuous embeddings for retrieval and computationally intensive coupled interaction networks for localization, thereby resulting in severe bottlenecks. To resolve this challenge, we propose a novel framework named Contextual Hashing with Decoupled Localization (CHDL). First, to address the efficiency issue, we introduce a temporal contextual hashing mechanism. This module projects high-dimensional video-text features into compact binary codes while preventing feature collapse through entropy maximization, thereby enabling ultra-fast retrieval through hardwarefriendly bit-wise Hamming distance computation. Second, we propose a decoupled multi-scale moment locator, that structurally decouples cross-modal semantic interactions from downstream temporal boundary prediction. By operating exclusively on one-dimensional relevance signals derived from a token-adaptive dynamic relevance model and using pyramidal dilated convolutions, this lightweight locator achieves precise temporal localization with minimal computational overhead. Furthermore, to ensure robust feature discriminability, we design a representation alignment objective that is based on dual-granularity mutual information maximization. This objective enforces alignment at both the global query-video level and the local token-frame level, thereby facilitating high-quality cross-modal interaction. The results of extensive experiments on the TVR and DiDeMo-FIG datasets demonstrate that CHDL achieves not only superior performance in localization accuracy but also a significant reduction in retrieval latency.