Panopticus is a groundbreaking omnidirectional 3D object detection system specifically designed for resource-constrained edge devices. It focuses on maximizing detection accuracy while adhering to tight latency constraints. The system features a configurable multi-branch architecture, where each branch leverages different modules suited for BEV-based 3D detection. By profiling runtime characteristics, Panopticus identifies and detaches heavy modules that exceed resource constraints. A spatial-adaptive execution scheduler further enhances performance by selecting the best image-branch combination in real-time, optimizing for anticipated accuracy and processing time according to the spatial context of incoming frames.
Panopticus introduces a novel omnidirectional 3D object detection system tailored for resource-constrained edge devices, optimizing detection accuracy while ensuring low latency.
The architecture of Panopticus is designed with a multi-branch structure that allows distinct modules for BEV-based 3D detection, accommodating varying inference needs across multiple camera views.
The spatial-adaptive execution scheduler within Panopticus plays a crucial role by selecting optimal branch-image combinations, enhancing system performance through real-time estimations of accuracy and inference time.
By profiling runtime characteristics and dynamically adjusting the execution based on inferred spatial distributions, Panopticus meets stringent latency and memory constraints effectively.
Collection
[
|
...
]