Part-in -whole matching

[bnjff]

Thanks for your work on this library! I got good results matching a model to individual objects extracted from the point cloud, however so far had no luck fitting the model to a more cluttered scene. Since the relevant part of the wiki is not available yet, could you please give a brief guidelines to make part-in-whole matching work?

[nmellado]

Hi @bnjff Yes, the part-in-whole documentation is still not available.

The main reason is that I could find robust-enough general guidelines working on multiple examples. This is not a limitation of the registration technique itself, but rather a limitation of the current implementation, which is too much tailored for the pairwise registration case.

After some time thinking about this problem, I believe the Super4PCS application needs a kind of expert mode, providing more parameters to the user. Presets could be added for the part-in-whole case. The main improvement would be to provide more control on the sampling and base extraction stages. I believe we need:

• a basis defined in the object cloud, with its size controlled wrt the object size
• same for the overlap, this parameter should be expressed wrt the object size
• an automatic estimation of the sampling parameters of the scene, wrt to the base size, etc.

This clearly needs a dedicated development effort. I can try to put this idea in the next release (after 1.1.3).

[bnjff]

To work with the current implementation, I ended up running object detection on RGB and using that to crop out the ROI point cloud with the object.

[nmellado]

Thanks for the feedback @bnjff . Could you share more details about you approach ? Is it fully automated ?Is it something we could add to the library ?

[aigerd]

guys, note that if the dataset in which you want to find an object is given offline and you have enoght space (e.g. O(n^2)), there is no point in using the (super)4PCS. in this case, one can index that data such that all the needed computations are done offline.

On Mon, Nov 13, 2017 at 6:20 PM, Nicolas Mellado notifications@github.com wrote:

Thanks for the feddback @bnjff https://github.com/bnjff . Could you share more details about you approach ? Is it fully automated ?Is it something we could add to the library ?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/nmellado/Super4PCS/issues/46#issuecomment-343971848, or mute the thread https://github.com/notifications/unsubscribe-auth/AGYw95j5naKk8WXSlDjav28Kr4uf0kG2ks5s2GxOgaJpZM4PIB5C .

[nmellado]

Thanks @aigerd , that's definitively true. BTW, it might be interesting to add such brute force approach in the library for offline processing and GT generation.

[Giessen]

Hi, is the "Part-in -whole matching" supported or not? I tried to match part of a point clound with another but the result is random... tks

[Giessen]

I would like to match object shown in this picture. Even with 2000 points, the result is still random. Any suggestions, please?

[Giessen]

Obj files attached. Hex.zip