Forum

Author Topic: dense point clouds and depth filtering  (Read 9922 times)

Chironex

  • Newbie
  • *
  • Posts: 2
    • View Profile
dense point clouds and depth filtering
« on: December 20, 2014, 12:23:39 AM »
I'm using Photoscan 1.0.4. on a 64 bit PC running Windows 8.1 Enterprise. I have 2 GTX 780 cards, a Quadro K5000, and 128 GB of RAM.

I processed 84 images of a tropical rain forest canopy as a test to determine the best combination of density (lowest, low, medium, high and ultrahigh) and filtering (aggressive or mild). I expected to see differences in point density, and for aggressive filtering to retain fewer points than mild filtering. Instead, I found that there are (1) differences in the spatial distribution of points, with some of these analyses resulting in large areas that have no points at all (holes in the point cloud, see images below). I also found that (2) some objects get filtered out by one or the other filters (mild or aggressive). The attached images illustrate the problem.

Here's the breakdown on processing time and point density for each run.

Dense reconstruction (ultrahigh, mild): 6.99 hours, 82,435,352 points
Dense reconstruction (ultrahigh, aggressive): 7.18 hours, 111,084,053 points

Dense reconstruction (high, mild): 1.03 hours, 35,371,879 points
Dense reconstruction (high, aggressive): 1.05 hours, 39,441,465

Dense reconstruction (medium, mild): 0.20 hours, 10,654,843 points   
Dense reconstruction (medium, aggressive):   0.20   hours, 11,497,366 points

Dense reconstruction (low, mild): 0.05 hours, 2,892,327 points
Dense reconstruction (low, aggressive):   0.05   hours, 3,025,643 points

Dense reconstruction (lowest, mild)   0.02   hours, 746,817 points
Dense reconstruction (lowest, aggressive) 0.02 hours,    754,450 points

Therefore, my questions are, (1) why do higher density reconstructions result in large areas with no points at all? (2) Is there some way to turn depth filtering off entirely, and if so, would this retain all points for a given density of reconstruction?

stihl

  • Sr. Member
  • ****
  • Posts: 408
    • View Profile
Re: dense point clouds and depth filtering
« Reply #1 on: December 20, 2014, 05:23:39 AM »
Interesting results.

May I ask at what limit did you set the tie-points at for aligning?
Also what are the images resolution and how many tie points were found in each image on average?

Chironex

  • Newbie
  • *
  • Posts: 2
    • View Profile
Re: dense point clouds and depth filtering
« Reply #2 on: December 20, 2014, 08:53:00 AM »
stihl,

The camera was a 12 MP Canon Powershot, and the pixels are about 5 cm on the ground. The point limit at the Align Photos stage was left at the default 40,000. Is Projections on the Ground Control pane the number of tie points per image that you are asking about? The median is about 5000.

Marcel

  • Sr. Member
  • ****
  • Posts: 309
    • View Profile
Re: dense point clouds and depth filtering
« Reply #3 on: December 20, 2014, 09:54:25 PM »
At 'Medium' the image is used at half of the size, because the image is downscaled it becomes much sharper. It could be that because of this extra sharpness Photoscan does find useful points on Medium, while on Ultra (full sized image) it has trouble finding good points in the same area.

This is the only reason I can think of. (In the end, everything always comes down to image quality)







StevenF

  • Newbie
  • *
  • Posts: 39
    • View Profile
Re: dense point clouds and depth filtering
« Reply #4 on: December 22, 2014, 02:15:46 AM »
I'm having the exact same problem using images taken with a Z/I DMC-1. I don't notice any difference in image quality or image texture between areas that generate quality points and areas that have large gaps. What really baffles me is that 'aggresive' filtering produces more points with fewer gaps, while 'mild' filtering produces fewer points with larger gaps.  In the beta version (1.1) you can set depth filtering to disabled but it produces far too much noise to deal with.

My current solution to this problem is to fill the gaps by merging points from different quality settings. I generate dense point clouds with different quality settings and export each one to LAS. Then I merge the LAS files and thin with a grid which selects the highest point in a grid cell of a given size. I select the highest point because I'm interested in tree heights and the lower quality settings tend to smooth over tree tops.

Some other software such as Imagine Photogrammetry and SocetGXP perform hierarchical image matching and have the ability to keep points matched at lower pyramid levels. It would be nice if PhotoScan implemented a similar functionality. The point clouds generated by PhotoScan are pretty impressive for how fast it works though.