neeravbm wrote:What kind of models do you generate automatically from point cloud? Is it pipes using ClearEdge3D or is it more than pipes?
When I said modeling I should have clarified that I was referring to lower level surface modeling, rather than object modeling. If we have different levels of data processing, e.g.:
Level 1: Measurements (e.g. point clouds)
Level 2: Surface (e.g. 3D polygonal mesh, terrain model)
Level 3: Features (e.g. valve, pump, chair, etc.)
Software like ClearEdge attempts to go from Level 1 to Level 3, which is a perfectly valid approach, but Level 3 features don't necessarily provide the same functionality as a Level 2 surface model. Most of our work requires Level 2 surfaces, which means that it is highly sensitive to the quality of the point cloud. If you have a person walking through your scan ClearEdge doesn't care because that person doesn't look like a pipe. If we build a surface model from that scan though, that person will be part of the model, which is usually not desireable.
neeravbm wrote:Also what do you exactly expect the software to do in terms of classification? From what I understand, software will find locations of different objects in your point cloud. But how you want to classify that object, for e.g. call it a pump, or a valve, or a chair, is up to you, right?
I would like software that I can use to efficiently assign points to arbitrary classes. I see creating the actual model, whether that is a Level 2 surface or a Level 3 feature as a separate process (although in many cases tightly coupled to the classification step). Exactly what objects are interesting can and does vary on a day-to-day and project-by-project basis. Also, as alluded to in the person example above, in many cases classification is as much about identifying points that represent things I want to exclude from the model as it is about identifying the points representing objects I want to model.
To give more concrete examples, on Monday, I might be scanning a city street with cars parked along the curb. I'm only interested in the street and the surrounding buildings so I need a classifier that can identify cars so that they can be excluded from my model. On Tuesday, I'm scanning a car to reconstruct it, so now I need a classifier tuned to identify the types of noise common in car scans, e.g. specular reflections off the headlights, mixed pixels and bogus distance measurements around the edges of the windshield, etc. On Wednesday I'm scanning a golf course and need to locate sprinkler heads marked with pink flags, so now I need a classifier that can identify pink flags. Those are all very specific scenarios and I could easily come up with many more. I think most people will have their own equally unique requirements, which is why I'm more interested in good tools rather than inflexible pre-built solutions that can't adapt to changing and diverse user needs.
To be generally useful I think any classification software would need to be a hybrid system, with both good manual tools that help humans work efficiently and perform the tasks machines can't do and automatic tools that can provide efficiencies wherever possible. The manual tools probably need to come first, after all if you want to train a classifier to identify sprinklers, first you're going to need to identify 1000 sprinklers by hand so you have labeled training data to build your classifier with.