Hacker News Reader

Comment by pontifier | Hacker News Reader

The problem I see with all of these scanners is the lack of quality... Are they all based on the same feature tracking core? Why isn't anyone doing more innovative reconstruction like tracking edge contours, or something else... even with a depth sensor most of the scanned models I've seen are crap.

View on HN

Cos its hard, thats why.

You're assuming that CCDs are perfect image sensors that don't have noise to cause feature trackers to jitter.

You're also assuming that sparsely featured objects only need simple back/belief propagation to make a good model, finally doing it in less than ten hours on an iphone at anything other than <320voxels^3 is pretty impossible.

Even decent commercial laser scanners only have limited resolution at this scale. Your best bet is either lightfield capture or http://web.media.mit.edu/~achoo/polar3D/camready/manuscript_... (which I've not read fully yet, however looks pretty sexy, even if its not very general.)

Most structure from motion or monocular SLAM algorithms use some kind of feature tracking and then estimate a dense reconstruction after they optimize the locations of all the cameras and all the features.

But it's not the only way. There's one notable alternative method, which is LSD-SLAM: https://github.com/tum-vision/lsd_slam

In general, it's extremely hard to get good quality models from a cellphone camera. Believe me, thousands of innovative researchers are trying their best to make it better.

The critical property of all similar systems is the way how you use it. Modern research systems, both based on depth sensors or cameras can in fact provide very good results if operated by a skilled person in a controlled laboratory environment (no shaky movement, proper light, no reflections, shadows etc.). This is a very different setting to using the system by a wide audience in the variety of all possible situations which might occur every day. One of our goals is to bridge this gap and make a system which works well for as many situations as possible and can be enjoyed by as many people as possible. Still a very hard task to do. This will happen gradually, one release at a time. Therefore, we will very appreciate your feedback on using the app - both when it works and when it does not. This will help us to improve the app.

Don't quite agree, most recent works produce 3D models of very decent quality. Have a look at this https://www.youtube.com/watch?v=XySrhZpODYs for example, but there are others based on volumetric data-structures that give fine results as well.

They are, the output you're so disdainful of reflects methods even more complex than you're imagining. It's a tough problem.

The best work I've seen recently (from MIT) uses two sensors with different polarization filters to get an additional signal to help with noise rejection and surface estimation.

Reply on news.ycombinator.com