Monitoring changes of 3D building elements from unordered photo collections

This paper proposes an automated method for monitoring changes of 3D building elements using unordered photo collections as well as Building Information Models (BIM) that pertain information about geometry and relationships of elements. This task is particularly challenging as existing construction photographs are taken under different lighting and viewpoint conditions, are uncalibrated and extremely cluttered by equipment and people. Given a set of these images, our system first uses structure-from-motion, multi-view stereo, and voxel coloring algorithms to calibrate cameras, reconstruct the construction scene, quantize the scene into voxels and traverse and label the voxels for observed occupancy. The BIM is subsequently fused into the observed scene by a registration-step and voxels are traversed and labeled for expected visibility. Next, a machine learning scheme built upon a Bayesian model is proposed that automatically detects and tracks building elements in presence of occlusions. Finally, the system enables the expected and reconstructed elements to be explored with an interactive, image-based 3D viewer where construction progress deviations are automatically color-coded over the BIM. We present promising results on several challenging construction photo collections under different lighting conditions and sever occlusions.