Automated image pre-processing and evaluation pipeline for improving single-image-to-3D generation with TripoSR.
This project investigates whether automated image pre-processing can improve the quality and consistency of single-image-to-3D reconstruction without modifying the underlying 3D generation model.
The pipeline applies image processing operations before passing the input image to TripoSR, followed by quantitative evaluation of the generated 3D meshes.
This project was developed as part of my CCS5990 Master's research project at Universiti Putra Malaysia.
Input Image
↓
Background Removal
↓
Object-Centered Cropping
↓
Adaptive Padding / Object Scaling
↓
Image Enhancement
↓
TripoSR
↓
3D Mesh
↓
Evaluation
- Background removal
- Object-centered cropping
- Adaptive padding and object-scale normalization
- CLAHE-based contrast enhancement
- Sharpening and edge-preserving processing
- TripoSR-ready image preparation
Additional Pipeline V2 experiments evaluate different object-to-canvas ratios and enhancement configurations.
Python · OpenCV · NumPy · Pillow · SciPy · Matplotlib · Trimesh · rembg · TripoSR · PyTorch · Blender
The controlled experiment currently evaluates three object categories:
Mouse · Bottle · Shoe
Each object is evaluated from five viewpoints:
Front · Back · Left · Right · Top
The experiments compare baseline and pre-processed inputs using geometric, cross-view consistency, and runtime measurements.
| Mouse | Bottle | Shoe |
|---|---|---|
 |
 |
 |
| Mouse | Bottle | Shoe |
|---|---|---|
 |
 |
 |
The generated meshes are evaluated using metrics including:
- Vertex and face counts
- Connected components
- Degenerate faces
- Watertightness
- Euler characteristic
Reconstructions generated from different viewpoints are compared to evaluate geometric consistency.
The project separately measures:
- Image pre-processing runtime
- TripoSR generation runtime
- End-to-end runtime
Ablation experiments evaluate how individual pre-processing stages affect reconstruction results.
Single-Image-to-3D-Preprocessing/
│
├── README.md
├── requirements.txt
├── .gitignore
│
├── scripts/
│ ├── preprocessing/
│ ├── corrected_pipeline/
│ └── pipeline_v2/
│
├── examples/
│ ├── inputs/
│ └── preprocessed/
│
└── results/
├── mesh/
├── view_consistency/
├── runtime/
├── ablation/
├── pipeline_v2/
└── figures/
scripts/preprocessing/ contains the core image pre-processing implementation.
scripts/corrected_pipeline/ contains the corrected controlled experimental pipeline, including 3D generation, mesh evaluation, view-consistency analysis, runtime benchmarking, and ablation experiments.
scripts/pipeline_v2/ contains subsequent pipeline optimization experiments, including object-to-canvas ratio selection, edge cleaning, foreground processing, CLAHE, and sharpening experiments.
Install the Python dependencies with:
pip install -r requirements.txtTripoSR should be installed separately and placed in the project environment according to its original installation instructions.
This repository contains the controlled experimental pipeline and selected evaluation results.
The initial controlled dataset is intentionally small and is primarily used to validate the experimental workflow. Expansion to additional object categories and public datasets is part of the ongoing research.
The current results should therefore be interpreted as experimental findings rather than as evidence of general improvement across all single-image-to-3D reconstruction tasks.
The goal of this project is to investigate whether a lightweight automated image pre-processing module can improve the reliability of single-image-to-3D reconstruction while keeping the underlying 3D generation model unchanged.
Kumbacat-leon
Master's Project in Computer Science
Universiti Putra Malaysia