賴冠穎 Kuan-Ying Lai

HarDNet-DFUS — Diabetic Foot Ulcers Segmentation for DFCU 2022. 

• I have led a team of 5 labmates to transform the HarDNet-MSEG into an accuracy-oriented version. I Proposed Transformer, Boundary Lose, Deep Supervision, and most methods in the work. It increases a 4% higher mean dice. 

OS Lab w/ NachOS: 

• The lab started by implementing the system call to understand how to handle the exception. 

• Implement the translation to map the address space to the physical memory to complete a correct context switch and achieve multi-programming. 

• Implement a Multilevel Feedback Queue w/ preemptive SJF, Non-preemptive priority, and RR. 

• Implement the linked index allocator to enlarge the file size limit 8 times. The related works contain modifying the file header, operation, and destructor to remove the instance correctly.

Multi-Object Tracking

• I visualize and observe the relationship between an object's prediction and confidence score. Then, based on my hypothesis, I filtered the input of Kalman Filter in ByteTrack to increase the ID F1-score by 1.0%.

Semantic Segmentation

• Inspired by super-resolution. I surveyed some super-resolution techniques to help upsampling in semantic segmentation. Then propose a method that assembles the SegFormer and Pixel Shuffle to speed up the model by 10% while maintaining similar accuracy.

Computer Graphics

• The framework is based on OpenGL and Matrix4. Implement Transformation and Lighting. The matrix values depend on the cursor, which allows us to control the model. Then uniform these matrices to the shader and render the model.