The NO FREE LUNCH theorem for point cloud processing

As scanning technology gets cheaper the availability of point clouds for buildings is increasing significantly. At the same time decades of research exist that has tried to convert point clouds to semantically rich Building Information Models, a practice that has been recently termed Scan2BIM. Despite the significant past research a breakthrough is not visible that allows us to convert point cloud data to general purpose BIM models. Since quite some time, I am therefore wondering whether a general purpose Scan2BIM conversion is possible at all. To me it rather seems as if conversion processes need to be closely steered by very detailed and specific information requirements. These information requirements should be based on a sound analysis of engineering decisions that are to be made on the information. Once it is clear what information to extract and in what detail this information is required, dedicated extraction algorithms can be developed. Looking at the recently published studies research seems to shift towards such specific purposes. However, such processes can hardly be labeled general purpose Scan2BIM.

The entire discussion reminds me of a paper that I was writing some years ago with Robert Amor and Bill East about the sense and non-sense of general purpose information models. While writing, Bill suggested that we should argue for a FREE LUNCH THEOREM (NFL) for information models. We all liked the idea, but the reviewers did not, so the NFL for information models never made it into the final publication. Bill’s idea was inspired by the NFL theorem in search and optimization. Once this theorem was established, it immediately stopped the extensive research efforts into the ideal general purpose optimization method. More about the NFL for search and optimization here.

Now years later I think we should consider a NFL for point cloud processing as well. For research the existence of such a NFL would have quite some ramifications. It would require a much more humble approach to point cloud processing focusing on very small purposeful engineering applications and the development of clear ontologies describing the knowledge required for these applications. These ontologies then need to steer the development of the geometric point cloud extraction methods. Developed methods would, however, not work for generalized purposes.

Leave a Reply