Projects
3D hand tracking and application to virtual characters
[github] [document] [video]
This project is part of the HKU Virtual Classroom Project, which uses multiple cameras to capture facial expressions and body movements to build a 3D virtual classroom.
The application used a single camera to capture hand movements and applied them to the virtual character. Improved 3D hand tracking algorithm; Leveraged deep learning and kinematics methods to calculate hand joint positions and angles; Deployed the algorithm in a Unity application using C++/C# and MediaPipe.
Smargo: An efficient and highly accurate solver for tsumego
[github] [document] [video]
Smargo is a python-based tsumego problems solver. It apply the Monte Carlo tree search algorithm to solve tsumego games, and improve the traditional Monte Carlo tree search structure to make it more suitable for determining and solving tsumego games.
Smargo also has its own dataset: Smargo dataset, which solves the problems of insufficient amount of tsumego data and inconsistent standards.
IndoorHIIT Motion Recognition
IndoorHIIT Motion Recognition is a smart AI program which can recognize users' fitness movements. It collected IndoorHIIT motion data from 50 testers, and applied Random Forest Classifier for motion recognition, utilized Wave Detection Method to count the number of movements.
This project also has the interface of the WeChat mini-program, deployed the model in the server and used terminal cloud architecture to achieve motion recognition.
Object Detection and Image Classification Using Raspberry Pi
This project is a deep learning related project. We selected MobileNetV2 and SSD models and trained the models for image classification and object detection, respectively. We deployed the trained models on Raspberry Pi to do run deep learning algorithms on this mobile.
We have made the project open source with a full tutorial document, which can be viewed at the link above.
Deep Forest with PaddlePaddle and Performance Evaluation
Deep forest is a cascade forest classifier model proposed by Zhihua Zhou, which is an integration of the traditional forest model in terms of width and depth.
This project uses Python + PaddlePaddle to implement deep forest. And we compare it with traditional machine learning algorithms (such as random forest, etc.).