High-performance video reasoning segmentation method based on time sequence marking
By using a multimodal large language model to generate timing marks in video inference segmentation, and performing timing dynamic aggregation and keyframe selection, the problems of insufficient timing context information capture and inaccurate keyframe positioning in the prior art are solved, and higher video inference segmentation performance is achieved.
Patent Information
- Application Number
- CN202510174580.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-18
- Publication Date
- 2025-06-06
AI Technical Summary
Existing video inference segmentation methods are difficult to effectively capture the rich timing context information in videos, and the keyframe positioning is not accurate enough, which limits the application of the model in downstream tasks.
A high-performance video inference segmentation method based on timing marking is proposed, using a multimodal large language model to generate timing markings, and through timing dynamic aggregation and keyframe selection strategies based on timing marking, the model's perception of spatial features and temporal dynamics is improved.
Through the encoding and dynamic aggregation of timing marks, the model's perception of video timing information is significantly improved, and the accuracy of keyframe positioning is improved, thus performing better than the previous method in multiple video and image inference segmentation tasks.