[Paper] Small traffic-sign detection using Alpha-Blending Feature Pyramid Network

This is my proposed paper of master's degree. (2020.02)

> http://www.dcollection.net/handler/yonsei/000000522129

 

- Abstract

The small object detection is an important sub-task in detecting and classifying objects of various sizes in scene. Especially, it is a challenging task in detecting small object such as traffic-sign in computer vision field. Beginning with the CNN model, previous work achieved outstanding performance using with deep learning. The benchmark for small traffic-sign detection was announced and the research have been continued. However, in order to improve detection performance of small traffic-sign, additional research is needed. In this paper, we propose a Alpha-Blending FPN model based on FPN(Feature Pyramid Network)[1] composed of ResNet-50[2]. To extract balanced feature map from the lateral connection of FPN, apply Alpha-Blending Element-wise operation. The network structure of Alpha-Blending FPN model based on FPN is varied to different models and the performance of each model is compared and evaluated. The learned model shows that the accuracy is improved for small size traffic-sign. We also deployed object detection service using Docker virtualization technology in the AWS cloud environment to actually test learned model.

 

- Keyword

Traffic-sign, Small object detection, Self-driving, Deep Learning, Neural Network, FPN, CNN, AWS, Docker, Alpha-Blending

 

- Paper Link

http://www.dcollection.net/handler/yonsei/000000522129

 

- GitHub

https://github.com/asyncbridge/tsinghua-tencent-100k

 

* Docker Based

https://github.com/asyncbridge/object-detector-tt100k-base-gpu

https://github.com/asyncbridge/object-detector-tt100k

https://github.com/asyncbridge/web-ui

 

- Test Web Site

http://ai.bakevision.com/task03