Training-free video corpus moment retrieval method based on self-adaptive calibration mechanism

By employing a training-free method with an adaptive calibration mechanism, and utilizing a large language model and a vision-language pre-trained network to construct cross-modal feature representations, this approach addresses the generalization capabilities and boundary localization accuracy issues of existing video corpus time-of-view retrieval methods in open-domain scenarios, achieving efficient and accurate video time-of-view retrieval.

CN122153119AActive Publication Date: 2026-06-05KUNMING UNIV OF SCI & TECH
5 Cites 0 Cited by

Patent Information

Application Number
CN202610623222.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-05-08
Publication Date
2026-06-05
Estimated Expiration
2046-05-08

AI Technical Summary

Technical Problem

Existing video corpus time-of-view retrieval methods suffer from weak generalization ability, high training and deployment costs, and insufficient accuracy in time boundary positioning in open domain scenarios. In particular, they lack in-depth analysis of text and video and dynamic boundary optimization when dealing with complex natural language queries.

Method used

A training-free method based on adaptive calibration mechanism is adopted. A large language model is used to parse natural language queries to generate structured query event chains. A cross-modal feature representation is constructed by combining a vision-language pre-trained network model. A mean-variance joint scoring mechanism and a gain-loss dynamic feedback strategy are introduced to achieve adaptive optimization of cross-modal semantic matching and time boundaries.

Benefits of technology

It improves cross-modal semantic alignment accuracy, reduces model deployment costs, enhances global semantic coherence, improves the robustness of precise temporal boundary positioning, and significantly improves the accuracy and robustness of retrieval results.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The invention relates to a training-free video corpus moment retrieval method based on an adaptive calibration mechanism, and belongs to the technical field of computer vision and multi-modal information processing. Comprising the steps of calculating an event-level similarity score based on a constructed query event chain and a constructed video event chain, and performing cross-modal retrieval in combination with a mean-variance joint scoring mechanism to obtain candidate proposals; in a time positioning stage, boundary-level association is established among video candidate proposals through a cooperation mechanism, self-adaption is executed by using a profit-loss dynamic feedback strategy, and time boundaries of the candidate proposals are iteratively optimized in a self-adaption manner. According to the method, a closed-loop retrieval process from text semantic analysis to video accurate matching is realized. Compared with the prior art, the method has the advantages that large-scale data labeling for training or parameter updating is not needed, the problems of generalization bottleneck and deployment cost of video retrieval in an open domain scene are effectively solved, and the precision and robustness of time positioning are remarkably improved under the zero training condition.
Need to check novelty before this filing date? Find Prior Art