SmartAdapt: Multi-branch Object Detection Framework for Videos on Mobiles

Researchers at Purdue University have developed SmartAdapt, a video object detection framework that identifies multiple execution branches and dynamically schedule which to run in real-time based on latency requirements and content-aware design. This technology can be easily integrated into existing object detection solutions (Faster R-CNN, EfficientDet, SSD, YOLO) previously thought to be too computationally expensive for mobile computing to allow them to perform well on mobile hardware. When running on an NVIDIA Jetson TX2, SmartAdapt was able to outperform all other softwares tested based on accuracy at varying levels of latency. Applications of this technology include mobile computer vision, autonomous vehicles, and object recognition/tracking.

Advantages

- Capable of running on mobile hardware (Jetson TX2 used in research)

- Easily integrated into state-of-art object detection platforms

- Lower latency and computationally lighter than current state of the art solutions

Applications

- Mobile computer vision

- Autonomous vehicles

- Object recognition/tracking

Technology Validation:

This technology has been validated by running the ILSVRC 2015 VID dataset. The technology outperformed content agonistic MBODF baseline, FastAdapt, by 20.9% to 23.6% mAP (mean average precision.).

Related Publications:

R. Xu et al., "Smartadapt: Multi-branch Object Detection Framework for Videos on Mobiles," 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 2518-2528, doi: 10.1109/CVPR52688.2022.00256.

TRL: 4

Intellectual Property:

Keywords: multi branch object detection,real time video analytics,integrated object detection frameworks,mobile computer vision,autonomous vehicle perception,low latency AI detection,edge AI acceleration,YOLO SSD Faster R CNN integration,efficient video recognition,mobile AI object tracking