Learning Compositions of Primitives for 3D Shape Representation
Zhishuo Liu
Individual Research Project (IRP) - Industry-Academia Collaboration
*Supervised by Dr. Hide Matsuki, Google (Feb 2025 – Present)
Developed a novel 3D primitive-based shape reconstruction system integrating superquadrics and signed distance field learning. Reproduced methods from papers such as SuperDec and used a transformer-based point cloud clustering approach to guide primitive fitting for improved 3D shape representation and reconstruction accuracy.

Efficient Mesh Reconstruction from 3D Gaussians with L1 Regularization and Pruning
Zhishuo Liu Supervised by Dr.Yizun Lin PDF / Code
We implemented L1 regularization and updated the pruning strategy in SuGaR, achieving an ~18% reduction in training time and lower memory usage while maintaining accuracy.

Aging Cause Analysis: Bayesian Analysis Based on Gompertz-Makeham Model
Zhishuo Liu
University of Birmingham Summer Research (Jul 2023 - Oct 2023)
Collected mortality data and patient information for major aging-related diseases in the United States, organized into 5-year interval format, and calculated conditional probabilities and related indicators. Applied Bayesian formulas to assess the impact of curing these diseases on human life expectancy, fitted the Gompertz-Makeham model, and determined constant mortality rate intervals for these diseases through hypothesis testing.

DFT Triangle: A Novel Spectrogram Feature for Stock Price Prediction by CNN
Zhishuo Liu, Zichen Wang, etc.
Abstract: Proposed a novel frequency-domain feature called DFT-triangle for stock price prediction, addressing the limitation of traditional DFT where similar price series exhibit distinct frequency features. The method stacks price series into 2D matrices with transferred frequency features, forming triangular patterns in scale-frequency coordinates. Evaluated on 470 stocks using CNN, DFT-triangle outperformed advanced sequential networks (Attention-LSTM, CNN-LSTM) and traditional econometric models (AR, ARIMA) in stock forecasting accuracy.
Code

2022 Contemporary Undergraduate Mathematical Contest in Modeling (CUMCM)
Zhishuo Liu, Zhiqi Ma, Jiayi Wu
National Second Prize
Abstract: Developed a physics-based dynamic model for wave energy capture devices using Newton’s laws and nonlinear differential equations. Implemented optimization algorithms to maximize energy conversion efficiency, achieving 163.8W output power with optimal damping coefficient. The system demonstrates stable periodic power output with 7% efficiency increase per 1% frequency rise.
Certificate / Paper (PDF) / Problem A

AI Fitness Coach: Real-time Exercise Analysis System
Xiaowen Zheng, Zhishuo Liu, Jia Lu, etc.
Silver Award – Internet+ Innovation Competition
A TensorFlow-based AI fitness coaching system that uses MediaPipe for real-time human pose estimation and machine learning algorithms (logistic regression, etc.) for exercise classification. Features include rep counting, form accuracy detection, and personalized exercise recommendations. The system analyzes body keypoints to provide real-time feedback on exercise performance and technique.
Certificate / Code

PPT2Script: AI-Powered Presentation to Speech Converter
Zhishuo Liu
A Flask-based web service that converts PowerPoint presentations into natural speech scripts using AI. The system parses PPT/PPTX files, extracts text content from slides, and generates coherent speech scripts through integration with Dify and ChatGPT (or other LLMs). Features include automatic slide parsing, customizable speech styles, and RESTful API endpoints for file upload and health monitoring.
Code / Demo