Sheng-Hsiang Hung

I’m an M.S. student in Computer Science at National Tsing Hua University (NTHU), advised by Prof. Hung-Kuo Chu.

Email  /  CV  /  Github  /  LinkedIn

Research

I'm interested in computer vision, deep learning, and image processing. Most of my research is about 3D reconstruction usually with radiance fields.

A load-balanced training pipeline for large-scale 3D Gaussian Splatting. By partitioning with a runtime-correlated load proxy and optimizing camera visibility, it cuts end-to-end training time by ~2× without sacrificing quality compared to SOTA methods.

Selected Projects

Built an image-to-3DGS/mesh reconstruction pipeline and a 60 FPS web viewer with distance/volume measurement tools achieving less than 1 m real-world error.

Tech: Javascript, React, Three.js, Python

Built an LLM-assisted JSON-to-CSG system with SDF primitives, boolean operations, GPU ray marching, interactive scene editing, and optional mesh extraction.

Tech: C/C++, OpenGL

Collaborated in a 4-person team to implement a real-time rendering engine with Blinn-Phong shading, deferred shading, shadow mapping, normal mapping, bloom, SSAO, SSR, FXAA, point light shadows, and area lighting.

Tech: C/C++, OpenGL

Collaborated in a five-person team to build a real-time interactive game with player controls, gameplay mechanics, scene logic, and visual feedback. Contributed to core gameplay implementation and integration to deliver a playable Unity-based project.

Tech: Unity

Led a two-person team to develop a CUDA/OpenMP parallel Ant Colony Optimization solver for the Traveling Salesman Problem on TSPLIB datasets. Optimized GPU-based next-city selection using parallel prefix sum, shared memory, bank-conflict avoidance, and reduced warp divergence, achieving up to 137x GPU speedup.

Tech: C/C++, CUDA, OpenMP, CPU/GPU profiling

Led a four-person team to build a U-Net audio super-resolution model with residual/skip connections and AFiLM-based feature modulation. Combined waveform and spectrogram losses, reducing LSD from 2.60 to 0.85 through two-phase training.

Tech: Python, PyTorch, U-Net, STFT, audio signal processing

Led a five-person team to develop a CNN-based speech emotion recognition system using a self-collected Taiwanese soap opera audio dataset. Extracted MFCC, Mel spectrogram, ZCR, Chroma STFT, and RMS features, and analyzed spectrogram differences between real-world drama audio and acted emotion datasets.

Tech: Python, Pytorch, Transfer Learning

Led a two-person team to build an FPGA-controlled handwriting plotter that converts keyboard-selected digits into coordinate-based stroke sequences. Implemented reusable Verilog FSM modules for line drawing, digit generation, X/Y stepper motor control, and PWM-based servo pen control.

Tech: Verilog, FPGA