Agisoft Metashape

Agisoft Metashape => General => Topic started by: b3059651 on January 12, 2015, 04:35:27 PM

Title: Photoscan process and publiced articles
Post by: b3059651 on January 12, 2015, 04:35:27 PM
Hi all,

I would like to follow a previous discussion http://www.agisoft.com/forum/index.php?topic=89.0 about the process in Photoscan. I found a nice recent article (Remondino, F., Spera, M.G., Nocerino, E., Menna, F. and Nex, F. (2014) 'State of the art in high density image matching', The Photogrammetric Record, 29(146), pp. 144-166.) where the authors explain how the different software work. I will present here some sentences of the article exactly as they are written and ask some questions in the end so to understand if this is the case of Photoscan of what they describe. 

...Nevertheless, from the autors' experience and from the achievable 3D measurement results, the implemented image-matching algorithm seems to be a stereo semi-global matching (SGM)-like method (for this study, version 0.9.0 Photoscan was used). Normally the software delivers results that are already meshed.....(pg. 151)

...PMVS employs a true-multi-image matching approach, meaning that for each object point visible in multiple images only one unique 3D point (which satisfies certain geometric conditions) is computed. On the other hand a 3D point cloud is computed for each pixel in the overlapping area of each of the stereopair in the Photoscan and SURE methods. In such cases , for n stereopairs, n 3D points corresponding to the same object point can be computed. This is particularly true in the case of large GSD and sub-pixel matching, leading to clusters of 3D points grouped near each other in the object space (but representing the same 3D point). This large number of points then can be successively averaged or statistically  reduced to a cloud of unique points, but the user needs to consider a proper workflow that takes into account the point-cloud processing requirements for point averaging, de-noising and filtering....... (pg. 161)

Questions:
1. From the first paragraph: does the resulting dense point cloud represent the raw data or the vertices of the triangulated mesh with the new version of Photoscan?
2. From the 2nd paragraph: Does we actually get many points near to each other representing the same object feature when we export the final point cloud after the optimisation part? Is the point average, flitering of de-noising process part of the in-build command 'construct the dense point cloud' or we, the users, should apply a seperate process to get the average points for each object from  the multiple image-pairs?

It would be ideal if you can confirm these statements because these can help us, the users, to understand in depth the concept and of course to reinforce the reasons why we get better accuracies with Photoscan than other software when it comes to explain that in a written report.

Thanks you so much for the time reading this and replying in advance.

Regards,