con-struct

Inspiration

The concept originated from a straightforward observation: real estate agents, architects, and interior designers spend numerous hours measuring and modeling spaces by hand. The process is laborious and time-consuming, whether it is for virtual staging, renovation planning, or building digital twin creation. "What if you could just take a photo and instantly get a usable 3D model?" we asked ourselves. We discovered that this was not only feasible, but also attainable within a hackathon timeframe, thanks to the development of AI-powered depth estimation and the availability of smartphone cameras. Our goal was to democratize 3D scanning, making it as simple as taking a photo.

What it does In less than 30 seconds, Wall Scanner AI converts a single wall photo into a 3D model that is ready for production.

Important Features: -Single-Image 3D Reconstruction: No additional hardware is needed  Modern neural networks are used in AI-Powered Depth Estimation  -Colored Point Clouds: Maintains color and texture details  -Wall Plane Detection: RANSAC automatically isolates the wall surface  -Multiple Export Formats - OBJ, PLY, STL, GLB for different use cases  -Real-Time Processing: Pipeline completion in a matter of seconds  -Web Interface: Use your browser to upload, process, and download files.

How we built it

Architecture Overview Our pipeline consists of five main stages: Input Image → Depth Estimation → Point Cloud → Mesh Generation → Export

Challenges we ran into

• Since we don't have actual camera calibration data, we use approximate values
• Camera Calibration Without Calibration Data-Different phones have different focal lengths, but we don't have calibration data.
• Scale Ambiguity -Monocular depth estimation only gives relative depth, not actual meters.
• Memory Issues with Large Images -Normal Estimation Failures

Accomplishments that we're proud of

Wall Scanner AI proves that modern AI can make professional 3D scanning accessible to anyone with a smartphone. What once required expensive LiDAR hardware and specialized software can now be done with a single photo and free open-source tools.

What we learned

-Modern neural networks are used in AI-Powered Depth Estimation  -Colored Point Clouds: Maintains color and texture details  -Wall Plane Detection: RANSAC automatically isolates the wall surface  -Multiple Export Formats - OBJ, PLY, STL, GLB for different use cases 

• Real-Time Processing: Pipeline completion in a matter of seconds  -Web Interface: Use your browser to upload, process, and download files. ## What's next for con-struct -Multi-image reconstruction for better accuracy -Automatic scale calibration -Texture mapping from original photo

Built With

• css3
• fastapi
• html5
• javascript
• open3d
• python
• pytorch
• transformers
• trimesh
• vanillajs