Abstract
Existing 3D perception methods can be computationally inefficient due to static grids that may process irrelevant spatial regions and may struggle to integrate driving context. A framework can use a global state vector, derived from data such as vehicle ego-state and environmental information, to manage a perception pipeline. The state vector may concurrently drive two mechanisms: a context-driven grid deformation that can create a dynamic, foveated voxel grid to concentrate resolution on pertinent areas, and a deep state injection method using adaptive instance normalization to condition a network’s feature extraction at multiple layers. This integrated system may improve computational efficiency and contextual relevance by dynamically allocating processing resources and adapting its internal strategy to a current driving scenario.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Kligvasser, Idan; Rivlin, Ehud; Cohen, Regev; Leifman, George; and Intrator, Yotam, "Context-Driven Grid Deformation and Deep State Injection for 3D Perception", Technical Disclosure Commons, (May 20, 2026)
https://www.tdcommons.org/dpubs_series/10196