update readme.md

README.md

@@ -1,90 +1,93 @@
----
-title: Multi-View 3D Reconstruction (MV3DR)
-emoji: 🎨
-colorFrom: gray
-colorTo: indigo
-sdk: gradio
-sdk_version: 4.44.0
-app_file: app.py
-pinned: false
-license: cc-by-nc-sa-4.0
----
-
 # Multi-View 3D Reconstruction (MV3DR)
 
-A web application for multi-view 3D object reconstruction using DUSt3R (Dense and Unconstrained Stereo 3D Reconstruction).
+Convert multiple 2D images into 3D models using DUSt3R (Dense and Unconstrained Stereo 3D Reconstruction). Simple web interface powered by Gradio.
+
+## Quick Start
 
-## Overview
+```bash
+git clone https://github.com/https-sharif/MV3DR.git
+cd MV3DR
+bash setup.sh
+python main.py
+```
 
-MV3DR is a production-ready system that generates 3D models from multiple 2D images through dense stereo reconstruction. The application features a dark-themed web interface built with Gradio, providing real-time visualization of depth maps, confidence heatmaps, and interactive 3D outputs.
+Open `http://0.0.0.0:7860` in your browser, upload 2+ images, and generate your 3D model.
 
 ## Features
 
-- Multi-view stereo reconstruction from 2+ images
-- Dual export modes: point cloud or textured mesh
-- Real-time depth map and confidence visualization
-- Advanced post-processing with background filtering
-- Point cloud cleaning and alignment optimization
-- Interactive 3D viewer with fullscreen support
-- Modular architecture with separated core modules
+- Multi-view 3D reconstruction from 2+ images
+- Export as point cloud or mesh (GLB format)
+- Real-time depth maps and confidence visualization
+- Background filtering and point cloud cleaning
+- Adjustable quality settings (100-1000 iterations)
+- Interactive 3D viewer with fullscreen mode
+
+## How It Works
 
-## Technical Stack
+1. **Preprocess**: Enhances image contrast and sharpness
+2. **Inference**: DUSt3R model predicts depth and generates 3D points
+3. **Alignment**: Merges multiple views into unified 3D space (300 iterations default)
+4. **Post-process**: Cleans point cloud and filters background
+5. **Export**: Generates GLB file for viewing/download
 
-- **Model**: DUSt3R (Dense and Unconstrained Stereo 3D Reconstruction)
-- **Framework**: PyTorch with mixed-precision inference
-- **Interface**: Gradio with custom dark theme
-- **3D Processing**: Trimesh for mesh/point cloud generation
-- **Visualization**: Matplotlib for heatmaps
+## Setup
 
-## Architecture
+**Requirements**: Python 3.10+, 16GB RAM, ~3GB storage
 
+Run the setup script:
+
+```bash
+bash setup.sh
 ```
-MV3DR/
-├── main.py              # Application entry point
-├── config.py            # Centralized configuration
-├── model.py             # DUSt3R model initialization
-├── pipeline.py          # Main reconstruction pipeline
-├── gradio_ui.py         # Web interface
-├── core/
-│   ├── inference.py     # Model inference with AMP
-│   ├── preprocessing.py # Image enhancement
-│   ├── postprocessing.py# 3D output generation
-│   └── visualization.py # Artifact generation
-└── assets/
-    ├── style.css        # Dark theme styling
-    └── fullscreen.js    # 3D viewer controls
+
+This installs PyTorch, Gradio, Trimesh, and downloads the DUSt3R model (~2.1GB).
+
+Then start the app:
+
+```bash
+python main.py
 ```
 
-## Pipeline
+### Manual Installation
+
+If the setup script fails:
+
+```bash
+python3 -m venv .venv
+source .venv/bin/activate
+
+pip install torch torchvision gradio trimesh scipy roma einops
+git clone -b dev --recursive https://github.com/camenduru/dust3r
+cd dust3r && pip install -e . && cd ..
+
+mkdir -p dust3r/checkpoints
+wget https://huggingface.co/camenduru/dust3r/resolve/main/DUSt3R_ViTLarge_BaseDecoder_512_dpt.pth \
+  -P dust3r/checkpoints/
+```
 
-1. **Preprocessing**: Contrast and sharpness enhancement
-2. **Pair Generation**: Create symmetrized image pairs
-3. **Inference**: DUSt3R stereo reconstruction with mixed precision
-4. **Alignment**: Multi-view point cloud optimization
-5. **Post-processing**: Depth cleaning and background filtering
-6. **Export**: GLB file generation (mesh or point cloud)
-7. **Visualization**: Depth maps and confidence heatmaps
+## Usage
 
-## Performance
+1. Upload 2+ images of your object from different angles
+2. (Optional) Adjust settings:
+   - Iterations: 100-1000 (default: 300)
+   - Point cloud or mesh export
+   - Background filtering
+3. Click "Run Inference"
+4. Download GLB file or view in browser
 
-- RTX 4090: ~15-20 seconds (3 images)
-- RTX 3080: ~25-30 seconds (3 images)
-- T4 GPU: ~40-50 seconds (3 images)
-- CPU: ~5-10 minutes (3 images)
+**Tips**: Use 4-8 images with good overlap between views. Matte objects work better than shiny ones.
 
-## Requirements
+## Configuration
 
-- Python 3.10+
-- CUDA-capable GPU (16GB+ VRAM recommended)
-- 16GB+ RAM
+Edit `config.py` to customize:
+- Image size (default: 512px)
+- Iterations (default: 300)
+- Server port (default: 7860)
 
 ## License
 
-Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
-Built on DUSt3R by Naver Corporation.
+CC BY-NC-SA 4.0 - Non-commercial use only
 
 ## Credits
 
-- DUSt3R: Naver Corporation
-- Gradio: Hugging Face
-- Trimesh: Michael Dawson-Haggerty
+Built with [DUSt3R](https://github.com/naver/dust3r), [Gradio](https://gradio.app/), [PyTorch](https://pytorch.org/), and [Trimesh](https://github.com/mikedh/trimesh).