Reconstructing 3D geometry from photographs is a classic Computer Vision problem that has occupied researchers for more than past few decades. Its applications range from robotics, 3D mapping, virtual world modeling, navigation to online shopping, 3D printing, computational photography, computer video games, or cultural heritage. However, these techniques are only matured enough recently to exit the laboratory controlled environment into the wild, and provide industrial scale robustness, accuracy and scalability.

Modern 3D reconstruction algorithms share the same basic processing pipeline: 1. Detect 2D features in every input imageand match among them. 2. Structure-from-motion algorithms to get the camera parameters and sparse 3D point sets. 3. Densification of the sparse point set to a dense point set. 4. Surface reconstruction algorihtms to get the final surface. The overal reconstrution pipeline above is usually refered as geometric approach. Researches have also proposed different ways to improve the final results such as photometric approach (photometric stereo and shape-from-shading) for detailed surface normals receovery which can help the final surface integration. Or better approach in texturing the mesh so that the final results look better even the underlying geometry is not perfect.

We have been working on various parts of the above pipeline. Below show some 3D recontruction results with our latest matching algorithm which is robust for wide baseline scenario.

Real World Reconstruction Results

1) HDB

2) HDB 2

3) Tank