Researchers propose a method for anytime computing in deep neural networks processing LiDAR point clouds, enabling dynamic input resolution scaling to balance execution latency and result utility. The approach allows models using pillars or voxels to adjust resolution at runtime without deploying multiple distinct models.
- The system uses a single DNN model that dynamically scales input resolution based on timing requirements.
- A deadline-aware scheduler predicts execution time for all possible resolutions at runtime to select the highest feasible resolution.
- Experimental results on the nuScenes dataset show significant performance improvements over existing anytime computing approaches.
- Deployment in a simulated autonomous driving system enables collision-free navigation while avoiding stalls caused by environmental complexity.
This method improves adaptability in cyber-physical systems by efficiently managing computational resources under dynamic operational constraints.