Update README.md

README.md

@@ -1,33 +1,57 @@
 ---
-title: Sparse Multi‑View 3D (Urban Planning)
-emoji: 🏙️
-colorFrom: gray
-colorTo: blue
+title: "2D → 3D Reconstruction (GLPN + Open3D)"
+emoji: 🏠
+colorFrom: indigo
+colorTo: purple
 sdk: gradio
-sdk_version: 4.44.0
+sdk_version: 4.29.0
 app_file: app.py
 pinned: false
+license: mit
+tags:
+  - depth-estimation
+  - monocular
+  - 3d-reconstruction
+  - open3d
+  - point-cloud
+  - mesh
+  - gradio
+  - huggingface
 ---
 
-Upload **3–30** images of an outdoor scene and get back:
-- **fused point cloud** (`fused.ply`)
-- **textured‑agnostic mesh** (`mesh.obj`) and **`mesh.glb`** for GIS/CAD viewers
+# 2D → 3D Reconstruction (GLPN + Open3D)
 
-### Capture Tips
-- Walk an arc with ~60–70% overlap between frames.
-- Mix wide + closer shots; include ground/verticals.
-- Avoid heavy motion (cars/people) if possible.
+This Space estimates **monocular depth** from a single RGB image using **GLPN**, builds an **RGB-D point cloud**, and reconstructs a **3D mesh** with Poisson surface reconstruction via **Open3D**.
 
-### Controls
-- **Matching mode**: `sequential` (good if you walked along a path), `spatial`, or `exhaustive` (slower, more robust for sparse unordered sets).
-- **Poisson depth / target triangles**: increase for more detail; decimate for lighter meshes.
+---
+
+## 🚀 How it works
+1. Upload an image.  
+2. GLPN (NYU pretrained) → predict relative depth.  
+3. Open3D → convert RGB + depth → point cloud.  
+4. Poisson reconstruction → mesh (downloadable `.obj` and `.ply`).  
+5. Preview depth map, mesh snapshot, and explore the mesh interactively.
+
+---
 
-### Outputs
-- Preview loads the **GLB**. Download the **OBJ** and **PLY** for further analysis.
+## 📦 Outputs
+- **Depth map** (colorized preview)  
+- **Point cloud (.ply)**  
+- **Mesh (.obj)** (with Gradio 3D viewer)  
+- **Mesh preview PNG** (best-effort offscreen render, if available)  
 
-### GIS / Urban workflows
-- Mesh is in **arbitrary scale**. Align in QGIS/ArcGIS/Blender using known distances or control points to metric scale.
-- For true georeferencing, add GCPs / scale bars in the scene and fit in external tools.
+---
+
+## ⚠️ Notes
+- Monocular depth has **no absolute scale** → geometry is up-to-scale only.  
+- For metric accuracy, swap in stereo, multi-view SfM, or metric depth models (ZoeDepth, Depth Anything v2).  
+- Works on **CPU or GPU** Spaces. GPU recommended for faster inference.  
+
+---
 
-### Compute
-- Works on **CPU** (slower). If your Space has CUDA, SIFT and PatchMatch will use GPU where possible.
+## 🛠️ Local Development
+```bash
+git clone <this-space>
+cd <this-space>
+pip install -r requirements.txt
+python app.py