WINTER CAMP


Due to the travel restrictions of covid-19, the consortium agreed to postpone summer school to winter, so it will be referred as “winter school” in this report. The winter school for 2022 academic year, hosted by UY, was held from Feb.3-16th, 2023.

23 participants, 5 teaching assistants, who are UY students, and 8 professors from the 4 universities attended this winter school. As have mentioned above, the winter school consists of 3 parts: lectures on advanced AI technologies, project on real problem solving, and academic-industry forum. 5 faculty members delivered lectures on deep-learning based signal processing, computer vision, 3D model representation and retrieval technologies and their application to agriculture, education and medical fields. Then 5 teams were built by mixing the students from 4 universities in one team to solve a given real problem using the leant knowledge. It’s notable that in each team a Japanese postgraduate student was assigned to support the PBL (Project-Based Learning) and acted as the student assistant. This kind of arrangement at one hand facilitates the newly established team to work on their topic more smoothly, at the other hand provide the chance for Japanese students to engage more in the program and deepen their communication with students from the other countries. 

In the Forum, keynote speakers Mr. Shihong LAO, (CEO, SenseTime Japan Ltd.) and Dr. Jonghoon Park, (CEO, Neuromeka Co., Ltd.) delivered speeches entitled as “AI powering future” and “Anatomy of A Commercial Collaborative Robot from the Control Viewpoint” respectively. The speakers shared their experiences in applying the AI technologies to the industry and also indicated the future direction of AI which widened students’ visions and helped them to understand the prospect of the AI industry and what kind of human resources are demanded.
In the Symposium, students from 5 teams presented their real-problem driven solutions with AI. The topics of their projects and research contents are as follows: (Team name ／ Topic)


・ DL3D／Deep Leaning for 3D Point Set Analysis
・ Grape Field Eye／Grape Leaf Disease Detection
・ Meat Lover／Yakiniku Recognition System: to Assist in CVD
・ Design Vision／Real Time Driver/Student Drowsiness Detection
・ Sound Sensor／Detection of Theft's Sound Using DNN

After the presentations, the excellent teams were selected according to the vote results. Team Grape Field Eye won the Gold Award and Team Meat Lover won the Silver Award. 

This event resulted in a deep understanding towards the latest direction of AI technology from the industry. In addition, the students got the opportunity to share and present their research projects, as well as exchange opinions with each other. It proved to be a fruitful forum. 

Insert Image

Insert Video

Help

Undoes the last command
Redoes the last command
Tab
Untab
Set a bold style
Set a italic style
Set a underline style
Set a strikethrough style
Clean a style
Set left align
Set center align
Set right align
Set full align
Toggle unordered list
Toggle ordered list
Outdent on current paragraph
Indent on current paragraph
Change current block's format as a paragraph(P tag)
Change current block's format as H1
Change current block's format as H2
Change current block's format as H3
Change current block's format as H4
Change current block's format as H5
Change current block's format as H6
Insert horizontal rule
Show Link Dialog

Summernote 0.8.2 · Project · Issues

Five teams’ projects 

Team A: DL3D／Deep Leaning for 3D Point Set Analysis

Team members: UniMAP 1, PKNU 1, UY 1, HDU 1

Research problem:
→Accurate classification of real-world 3D point sets by tackling the following challenges such as scarce labeled data and diverse corruptions inherent in real-world 3D point sets to improve accuracy.

Results achieved:
1. The accuracy of DNN model (PointNet: 30.3%) can be improved by using both data augmentation and pre-processing methods.
2. The best results in accuracies achieved by the supervised pretraining with data augmentation method (shear, severity:5), which is 42.86%.

Team B: Grape Field Eye/ Grape Leaf Disease Detection

Team members: UniMAP 2, PKNU 1, UY 1, HDU 1
 
Research problem:
→The early and correct detection and identification of the grape disease is greatly important in reducing the disease progression and economic cost.
→Need a very lightweight model allows the model to run on a smartphone alone. 

Results achieved:
1. Best model (~54 M parameters) (cloud server): Yolov8x, batch size= 1
(All class: 0.696 mAP50, Healthy: 0.76 mAP50, Disease: 0.613 mAP50)
2. Light weight model (~7 M parameters) (android): Yolov5s, batch size= 64
(All class: 0.699 mAP50, Healthy: 0758. mAP50, Disease: 0568. mAP50)

Team C: Meat Lover／Yakiniku Recognition System: to Assist in CVD

Team members: UniMAP 2, PKNU 1, UY 1, HDU 1

Team D: Design Vision／Real Time Driver/Student Drowsiness Detection

→Team members: UniMAP 1, PKNU 2, UY 1, HDU 1
 
Research problem:
→The risk of traffic accidents in drowsy drivers is estimated to be four to six times higher than awake drivers.
→65 percent of the students feel sleepy during the online classes between 1-3 times a week. There is no one to keep a track of the
students’ behavior during the conduction of course.
→ Drowsiness detection is necessary to avoid these problems
Methods: Head pose estimation + Eye closed detection
We used Media-pipe pre-trained model and created the GUI application.

How to use:
・Camera image is displayed on the left and 　
　detection results on the right.
・Elapsed time from looking down and from eyes
　closed are displayed in the lower right corner.
・Warning time can be changed.
・Warning sound can be changed.

Results achieved:
 1. Signs of drowsiness could be detected based on facial orientation and eye closure.

Team E: Sound Sensor／Detection of Theft's Sound Using DNN

Team members: UniMAP 1, PKNU 1, UY 2, HDU 1
 
Research problem:
→Use Microphone-based Surveillance System (MSS) to detect fruit thief
→Machine Learning on Audio Signal

Results achieved:
1. Precision 99%, Recall 91%
2. Implemented on Raspberry Pi4