XiYOLO: Energy-Aware Object Detection via Iterative Architecture Search and Scaling
A study published on arXiv presents XiYOLO, a framework for object detection that prioritizes energy efficiency, specifically tailored for diverse edge devices. This framework integrates an energy-conscious XiResOFA search space, a dual-phase energy estimator, and an iterative search process to pinpoint a singular energy-efficient base architecture. Subsequently, compound scaling adapts this foundational design into the XiYOLO family, accommodating various deployment budgets and facilitating clear accuracy-energy tradeoffs based on limited hardware measurements. Tests conducted on PascalVOC, COCO, and actual device deployments demonstrate that XiYOLO offers a superior energy-accuracy balance compared to YOLO baselines.
Key facts
- XiYOLO is an energy-aware object detection framework for edge devices.
- It uses an energy-aware XiResOFA search space, two-stage energy estimator, and iterative search.
- Compound scaling creates the XiYOLO family across deployment budgets.
- Experiments conducted on PascalVOC, COCO, and real-device deployment.
- XiYOLO achieves stronger energy-accuracy tradeoff than YOLO baselines.
- The paper is available on arXiv with ID 2605.06927.
- The framework addresses energy, latency, and memory constraints.
- It enables interpretable accuracy-energy tradeoffs under sparse hardware measurements.
Entities
Institutions
- arXiv