Adaptive-Cloud: Dynamic Computation Control for 3D Object Detection from LIDAR Point Clouds
Mir Sayeed Mohammad, Uday Kamal, Saibal Mukhopadhyay
AI summary
Problem
Static 3D object detection models waste computational resources on simple scenes, creating a bottleneck for real-time deployment on resource-constrained edge devices.
Approach
The method uses a shared feature extractor with parallel backbones of varying widths, guided by a feature gating module and a lightweight surrogate loss predictor to dynamically route scenes to the most efficient model at runtime.
Key results
- 41.4% average reduction in computational overhead (FLOPs)
- Only 2.44% drop in mean average precision across diverse driving scenes
- Real-time runtime adaptability without significant latency overhead
- Preserved downstream path planning safety and performance
Why it matters
Enables efficient, high-performance 3D perception for autonomous vehicles and edge robots by dynamically balancing accuracy and compute in real time.
Abstract
In this work, we introduce an adaptive hierarchi- cal framework for efficient 3D object detection from point cloud data,designedtodynamicallybalancecomputationalefficiencyand detection performance. Our approach employs a shared feature extractor and multiple detector backbones of varying widths, en- abling selective activation of models based on the complexity of the input scene. A novel feature gating mechanism dynamically de- termines the most relevant features for reduced-width backbones, while a surrogate loss prediction module ranks models in real-time, ensuring optimal backbone selection with minimal overhead. This adaptive strategy reduces compute costs by 41.4% while maintain- ing a negligible 2.44% reduction in detection accuracy across a range of real-world driving scenes (urban, highway, residential, campus, person) from the KITTI dataset. By addressing runtime adaptability—a critical gap in existing 3D detection frameworks— our method provides a significant algorithmic improvement for high-performance detection models in resource-constrained environments.