YOLO-LITE

A real-time object detection implementation of YOLO

About YOLO-LITE

YOLO-LITE is a web implementation of YOLOv2-tiny trained on MS COCO 2014 and PASCAL VOC 2007 + 2012.
What is Object Detection?

Object detection is a field in computer vision where the task is find and bound the location of certain objects in a given image. Having a low computation real time object detection algorithm allows virtually any device to be able to interact with its surroundings.

What is our goal with Yolo-Lite?

Our goal is to create an architecture that can do real-time object detection at a speed of 10 FPS and a mean average precision of about 30% on a computer with out a GPU.

What is our contribution?

We developed a yolo based architecture that can achieve 21 FPS on a Dell XPS 13' running on darkflow. This is 9x faster than the original tiny yolo v2. Our mean average precision is 33.57% compared to 40.48% when trained on VOC. Our model achieves its speed by shrinking the standard YOLOv2-tiny model and also getting rid of batch normalization. You can read more about our project here and find our code here. We have a live demo of our architecture trained on VOC and COCO datasets. Both of these demos run completly on the users computer. The FPS is lower than what would be achieved running on a local computer.

This project was created through the Data Mining REU at University of North Carolina Wilmington. NSF grant DMS-1659288

YOLO-LITE detection of us (missed the bottle and cups in our hands).

Dataset mAP FPS
VOC 33.57% 21
COCO --- ---
Examples
From our trained coco network

GET TO KNOW US

Jonathan Pedoeem

Research Intern

Jonathan Pedoeem

Rachel Huang

Research Intern

Rachel Huang

Dr. Cuixian Chen

Principal Investigator

Dr. Cuixian Chen

Thai Thompson

Graduate Assistant

Summerlin Thompson

Dr. Yishi Wang

Principal Investigator

Dr. Yishi Wang