MINGJIAN LI

NYU Video Lab: 3D Dancing Human Reconstruction Dataset

A benchmark dataset with multiview, multiframe human dancer sequences, including building a full pipeline for raw data processing, 3D Gaussian construction with PSNR ≈ 34.

Unity player integrated with AR/VR capabilities for real-time 3D Gaussian Splatting visualization, enabling interactive playback on mobile devices with seamless cross-platform deployment.

Unreal Engine plugin with native VR support for immersive 3D reconstruction visualization, featuring high-performance rendering and interactive VR display capabilities for enhanced spatial understanding.

Deep Learning Based Accelerated MR Image Reconstruction

GitHub RepositoryResearch ReportPresentation

Deep LearningMRIPyTorchCUDAFastMRIDiffusion Models

Research on deep learning–based magnetic resonance (MR) image reconstruction, aiming to accelerate scan times by reconstructing high-quality images from undersampled k-space data.

Comprehensive evaluation study of supervised, self-supervised, and zero-shot learning approaches using CNN, UNet, and diffusion-based models (Variational Networks, Score-based Diffusion Models) on the FastMRI dataset.

Technical Achievements & Results

• ▹VarNet Implementation: Achieved PSNR of 34.62 ± 3.78 dB with 4× acceleration and 0.24s inference time per slice
• ▹Score-Based Diffusion: Implemented SENSE parallel imaging achieving 12.5× speedup (from 30,000 to 2,000 evaluations)
• ▹Multi-GPU HPC: Refined PyTorch/CUDA implementations with distributed training achieving 4× speedup in performance
• ▹Systematic Evaluation: Tested 7 checkpoints to identify optimal model (epoch 50) with best generalization performance

Image Reconstruction Pipeline

Undersampled

Low-quality input

Deep Learning Model

UNet / Diffusion

Recovered

High-quality output