Single-image tree structure reconstruction using transformers and L-system

Single-Image Tree Structure Reconstruction Using Transformers and L-System

This project investigates the pairing of tree tokens encoded by L-systems with visual features from tree images to enable single-image tree reconstruction.

Lin Zhengyang

Master Student, NUS Computing

stock-vector-tree-big-data-technology-network-connection-on-blue-background-online-digital-cloud-storage-2458508255-16x9

Campus as a Living Lab (CALL@NUS)

Single-Image Tree Structure Reconstruction Using Transformers and L-System

This project investigates the pairing of tree tokens encoded by L-systems with visual features from tree images to enable single-image tree reconstruction.

Lin Zhengyang

Master Student, NUS Computing

This project investigates the pairing of tree tokens encoded by L-systems with visual features from tree images to enable single-image tree reconstruction.

The main innovation lies in applying the Vision-Language Pretraining (VLP) paradigm to pair L-system tokens, which represent hierarchical tree structures, with image features.

One of the unique contributions of this work is the application of VLP techniques to structured tree data, bridging multimodal learning across visual and hierarchical data.

Additionally, the project evaluates reconstruction across different tree species—Acacia, Birch, Maple, and Pine—highlighting species-specific challenges in the reconstruction process.

About the author

I am a Master's student from the School of Computing (SoC), majoring in Computer Science with a specialization in Artificial Intelligence (AI).

My primary interests lie in the rapidly developing fields of machine learning, AIGC (Artificial Intelligence Generated Content), and 3D computer vision.

What drew me to participate in this project was my interest in these rapidly evolving domains. The opportunity to explore and apply machine learning and 3D computer vision techniques to complex real-world problems, such as tree reconstruction for urban modeling, was both fascinating and intellectually stimulating.

Through this project, I gained valuable experience in integrating visual data with hierarchical structural representations, specifically using L-systems within a Vision-Language Pretraining framework.

I developed insights into multimodal learning and addressed challenges related to tree reconstruction.

Additionally, I learned how to manage structural variability across different species in the context of urban digital twin applications.

More about CALL@NUS

Home
Home
Research Capabilities
Campus as a Living Lab
Single-image tree structure reconstruction using transformers and L-system

Sidebar

Main Menu

About the author