3D Reconstruction from Public Photos with Machine Learning
2 days ago
- #Computer Vision
- #Machine Learning
- #3D Reconstruction
- The article explores the possibility of reconstructing 3D models from public photos using machine learning.
- Public photos from sources like Reddit, YouTube, and Google Maps can be utilized for 3D mapping.
- Cameras remove 3D information (depth) when taking photos, but ML can help recover it.
- Key components needed for 3D reconstruction: depth at each pixel (Z) and the camera's focal length (f).
- Apple's DepthPro model provides metric-scaled depth and estimates focal length, enabling accurate 3D reconstructions.
- Examples of 3D reconstructions include COEX Mall, a forest, NYC Skyline, Safeway, and Singapore Airport.
- The DepthPro model struggles with large-scale scenes like the NYC Skyline due to training dataset limitations.
- The process involves estimating depth masks, mapping pixels back to 3D, and visualizing point clouds with Open3D.