About me

Recent Computer Engineering graduate from Hong Kong Univerisity of Science and Technology (HKUST), seeking to use computer vision and machine learning experience in an entry-level position. Posses with both full-time and internship experiencing building and testing machine learning algorithms in Python and data analysis.

Work Experience

Computer Vision Researcher at Dayta AI Ltd.

• Assisted in building object tracking and detection for videos.
  
• Implemented uniform and action classification on shop surveillance videos.
  
• Optimized demographic classification.
  

MOKE image processing Engineer Intern at HKUST

• Collected Magneto-optic Kerr effect (MOKE) images from electronic microscope under suitable settings.
  
• Optimized the quality of skyrmion images using different computer vision algorithms from OpenCV and ski-image.
  
• Designed user-friendly Graphic User Interface by QML.
  

Research Experience

Research Assistant at Spintronics Quantum Material Laboratory

• Collected highly variable driving data for visual navigation.
  
• Conducted performance analysis on different deep neural networks with robotics car and evaluated the result on autonomous racing.
  

Undergraduates Researcher in Undergraduates Research Opportunities Program

• Conducted experiments on magnetic memory and memristor.
  
• Implemented an automatic testing system among magnetic field sensors, Keithley source meters and lock-in amplfier.
  
• Implemented real-time plotting for efficient data analysis.
  

Projects

CT Image Segmentation

• Performed segmentation on 3D CT tumour images by UNET with residual connection.
  
• Applied 3D image augmentation to increase data variety.
  
• Implement 3D dice loss to reflect the similarity between the ground-truth label and prediction.
  
• Achieved 0.83 on dice, 0.67 on Jaccard, 5.26 on average surface distance and 55.18 on 95% Harsdorf Distance.
  

Surgery Video Phase Recognition

• Implemented a temporal recognition network to classify surgical phase recognition.
  
• Used ResNet50 to extract features for each frame and LSTM to capture temporal features among 3 nearby frames.
  
• Applied data sampling to solve data imbalance.
  
• Achieved training accuracy of 97% and testing accuracy of 98%.
  

Self-driving car on Carla with CNN

• Implemented a temporal recognition network to classify surgical phase recognition.
  
• Used ResNet50 to extract features for each frame and LSTM to capture temporal features among 3 nearby frames.
  
• Applied data sampling to solve data imbalance.
  
• Achieved training accuracy of 97% and testing accuracy of 98%.