CHINA LITERATURE — AI RESEARCH
TTS & Generative Systems · College 1st Year
Multi-emotion TTS and generative imaging for production workflows.
- Led multi-emotion TTS framework development; compared LoRA vs VC for quality and control.
- Trained 4 bespoke emotion LoRAs; measured timbre consistency and intensity.
- Optimized inference pipeline, reducing model load time from 8.91s to 3.62s.
SJTU — 3D RECONSTRUCTION
NeRF & Gaussian Splatting · College 1st Year
Generalizable reconstruction across NeRF and 3D/4D GS.
- Experimented with baseline NeRF and 3D Gaussian Splatting methods.
- Surveyed key novel view synthesis methods from Mip-NeRF to Zip-NeRF and 4D GS.
- Assisted research on generalizable 3DGS inspired by MVSGaussian and PixelSplat.
- Designed a robust MVS + SfM fusion for stronger point-cloud initialization.
- Investigated editability challenges in 3DGS and evaluated practical solutions.
FALL DETECTION — SINGLE RGB
Classical ML & CNN · Grade 11
Comparing classical models and CNNs for RGB-only fall cues.
- Benchmarked DT, RF, SVM, MLP, and a custom CNN on curated video clips.
- Engineered motion-based features for classical baselines.
- Assessed accuracy vs. computational cost for practical application.
INSTRUMENT DIGIT READER
OCR + CNN Pipeline · Grade 12
Camera-to-classifier pipeline for lab instrument displays.
- Built a reusable OCR preprocessing pipeline coupled with a CNN digit classifier.
- Pre-trained on a synthetic dataset then fine-tuned on real-world images.
- Integrated camera capture hardware directly with the classification software.