Project Web-based Object Detection Service

[업데이트 2019.05.08 16:21]

 

This is a web-based object detection service. It provides detecting various type of objects. The state-of-the-art deep learning computer vision technology, “YOLO” based on Darknet is used. The Darknet is CUDA and C++ based deep learning framework. The pre-trained YOLOv3-608 model which is trained by the COCO dataset is deployed. I created new dataset for crested gecko detection. It is created based on Pascal VOC format. To train the new dataset, deploy training container module. It is connected to AWS S3 and send/receive the files between the training container module and S3. The Caffe deep learning framework is also used for my research in my paper. Tsinghua-Tencent 100K dataset for China traffic-sign detection is used for training. This is also deployed by Docker container.

 

•Developed web UI and object detector micro-service based on Docker container. The development environment is as follows.

- Used python language for implementing the container modules.

- Implemented Docker container module for the object detection. By including all of the python packages and the other development dependencies in the Docker container module, it makes the deployment easy.

- Implemented REST API interface of the object detection module based on Flask.

- Deployed Darknet deep learning framework and used GPU computing for the object detection.

- Deployed pre-trained YOLOv3-608 model which is trained by the COCO dataset. The COCO dataset consists of 80 classes.

- Deployed the YOLOv3-608 model which is trained by the crested gecko dataset.

- Deployed Caffe deep learning framework and used GPU computing for the object detection. Deployed the FPN-CNN model which is trained by Tsinghua-Tencent 100K dataset. The FPN-CNN model is a proposed model in my paper.

 

•Project Period: 2019.01-12

•Web URL : http://ai.bakevision.com

 

* The detection services can be unloaded due to the deep learning training task for my paper research. My GPU memory 11GB and it is not enough to deploy all of services. :)