Patt PHURTIVILAI

Hello! I’m Patt Phurtivilai, a research postgraduate in Artificial Intelligence at the University of Hong Kong. I conduct 3D vision research in CGVU Lab under the supervision of Prof. Taku Komura.

I am passionate about building AI that understands and interacts with the real 3D world. My research focuses on computer vision and graphics, including 3D tracking, reconstruction, and neural rendering.

My current work explores self-supervised multi-view tracking and 3D perception systems developed through collaborations with academic and industry researchers.

   /      /   pattgene [at] connect (dot) hku (dot) hk

Download CV

TrackFish3D: Self-Supervised 3D Tracking of Schooling Fish from Multi-view Videos
Jun 2025 – Present
First Author Paper — Proceeding in NeurIPS 2026

Achieved dense 3D multi-object tracking of visually identical fish without any manual annotations by developing a fully self-supervised framework that learns cross-view correspondence and temporal identity linking from geometric consistency across calibrated multi-view video.

EndoX: A GPU-Accelerated Endoscopic Perception Simulation Framework
Jun 2025 – Present
Proceeding to MICCAI 2026 — in collaboration with NVIDIA AI

Established 3DGS rendering benchmarks for small intestine datasets by contributing a 3D Gaussian Splatting model and conducting systematic evaluations within the simulation framework.

PHURTIVILAI PATT, Huang Leyang, Yinqiang Zhang, Lei Yang,The University of Hong Kong
DensifyBeforehand: LiDAR-assisted Content-aware Densification for Efficient and Quality 3D Gaussian Splatting
First Author Paper — Accepted Technical Brief, CVM 2026

This paper addresses limitations in 3D Gaussian Splatting (3DGS) methods, specifically their dependency on adaptive density control, which can lead to artifacts and inefficiencies. We propose a "densify beforehand" approach that combines sparse LiDAR data with monocular depth estimation to enhance scene initialization. Our ROI-aware sampling prioritizes important regions, creating a dense point cloud that improves visual fidelity and efficiency. This method reduces redundancy in Gaussians, allowing for better quality with lower resource consumption and training time, validated through comparisons on four new datasets.

[Page] [Paper]

Patt Phurtivilai, Lio Qing, Ryan Chan The University of Hong Kong
3D Digital Twins for MakerSpace
Presented at 9th International Symposium on Academic Makerspaces (ISAM2025)

Leyang Huang, Patt Phurtivilai, Xiang Chen, Syed Nafis Ishrak, Jose Alberto Espino Pitti, The University of Hong Kong
Fostering AI Makerspace: A Journey through Introductory Hands-On Workshop
Presented at 8th International Symposium on Academic Makerspaces (ISAM2024)

InnoGrow: Non-Toxic Weed Killing Box

• Developed INNOGROW: a non-toxic laser weed control solution for sustainable farming in Asia.
• Trained a weed detection model with a custom dataset; integrated with NCNN on Raspberry Pi for a compact design.
• Controlled laser shooting with a servo using an ESP32 board.

[Video] [Code]

Digital Twins: 3D Gaussain Splatting Reconstruction

• Collaborating with MTR, building a digital twins for tunnels.
• Lidar Scaning and processing.
• 3D Gaussian Splatting using Lidar Slam initialization.

Multimodal AI Assistant: RAG system for both image and text

• Developed a multimodal RAG system for processing and generating text and images.
• Created a data preprocessing pipeline with layout detection and Azure OCR for content extraction.
• Designed a retrieval mechanism for interleaving text and images, enhancing traceability and accessibility.

[Video] [Code]

Grounding-Dino: zero shot training with real time camrea feed
participate in HKU innoshow2024

• Developed a real-time object detection and tracking system from Grounding DINO and live feed.
• Implemented keyword-based detection for continuous object tracking.
• Showcased advanced computer vision techniques for dynamic mobile interaction.

[Video] [Code]

Azai Solution - CRM + Payment API Application
Winning the 1st prize for Cyberport CUPP 2023 micro funding

• Developed AZAI, a smart CRM app for landlords in Thailand using React Native and TypeScript.
• Integrated payment APIs and Firebase for efficient contract and property management.
• Utilized Next.js for backend development, adding analytics features to improve landlord-resident interactions.

[Page] [Video]