BaliMask3D
Optimizing Single-View 3D Reconstruction for Balinese Masks
An undergraduate thesis and research prototype evaluating deep learning-based reconstruction models and mesh optimization.
View project repository (opens in a new tab)Evaluated TripoSR, InstantMesh, and Wonder3D using 972 image samples and seven geometry and render-based metrics.
- Organization
- Undergraduate thesis and research project
- Role
- Research and model implementation
- Timeline
- Timeline not publicly specified
- Status
- Research prototype
- Contribution
- Implemented models, fine-tuned TripoSR and InstantMesh, optimized meshes, and evaluated results.
- 0127 Balinese masks
- 02972 image samples
- 03Seven evaluation metrics
Context
An undergraduate thesis on Balinese mask reconstruction
The research title is “Mesh Optimization in Deep Learning-Based Single-View 3D Reconstruction Models for Balinese Mask Objects.”
Role
Research and model implementation
My contribution covered model implementation, TripoSR and InstantMesh fine-tuning, mesh post-optimization, and evaluation. A public timeline is not stated because it has not been verified.
Challenge
Evaluating reconstruction and mesh optimization from a single view
The work examined deep learning-based single-view 3D reconstruction for Balinese mask objects using TripoSR, InstantMesh, and Wonder3D.
Contribution
From dataset to evaluated research prototype
- 01
Prepare the dataset
Use 27 Balinese masks with 36 views each, totaling 972 image samples.
- 02
Implement models
Implement TripoSR, InstantMesh, and Wonder3D.
- 03
Fine-tune and optimize
Fine-tune TripoSR and InstantMesh and apply mesh post-optimization.
- 04
Evaluate
Compare outputs with seven geometry and render-based metrics.
Verified outcome
A research prototype evaluated across three models
The dataset contained 27 Balinese masks, 36 views per mask, and 972 image samples. TripoSR, InstantMesh, and Wonder3D were evaluated using seven geometry and render-based metrics.
Project status
Research prototype with a conditional web demo
The project is a research prototype. Its web inference demo is not continuously public because the GPU service must be running.
Reflection
Connecting model research with usable evidence
What I learned
The project combined model implementation, fine-tuning, mesh post-optimization, and evaluation rather than treating reconstruction output as sufficient evidence by itself.
What I would do next
Future documentation should explain model outputs and evaluation evidence for non-technical readers without implying that the GPU-dependent demo is continuously available.