A spatiotemporal-aware key-value cache compression method for stream-oriented three-dimensional reconstruction

By using a spatiotemporally aware key-value caching compression method, the problems of cache size growth and reconstruction quality degradation in streaming 3D reconstruction are solved, achieving efficient memory management and improved inference efficiency, and is suitable for streaming 3D reconstruction.

CN122244310APending Publication Date: 2026-06-19UNIV OF SCI & TECH OF CHINA

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
UNIV OF SCI & TECH OF CHINA
Filing Date
2026-03-16
Publication Date
2026-06-19

Smart Images

  • Figure CN122244310A_ABST
    Figure CN122244310A_ABST
Patent Text Reader

Abstract

This invention relates to the field of streaming 3D reconstruction technology and discloses a spatiotemporally aware key-value cache compression method for streaming 3D reconstruction. The method includes: receiving a continuous stream of image frames, extracting features from each image frame and inputting them into a causal Transformer architecture for inference; maintaining a historical key-value cache during inference and performing spatiotemporally aware cache compression management, specifically including: time-aware short-term cache management, spatially-aware long-term cache management, and block-based multi-frame joint inference. This invention can significantly reduce memory consumption and accelerate the model inference process while ensuring 3D reconstruction quality and temporal consistency, and is suitable for online, real-time, and long-term streaming 3D reconstruction application scenarios.
Need to check novelty before this filing date? Find Prior Art