Export OpenSfM reconstructions to different geographic coordinate systems

[paulinus]

tracking here mapillary/OpenSfM#199

Ideally, any proj4 coordinate system should be supported

files to be exported:

• reconstruction.json
• reconstruction.nvm
• depthmaps/merged.ply

[paulinus]

Added a command to export the reconstruction and dense point cloud to coordinate systems defined by proj.4 lines.

The exported reconstructions though have very large numbers. Operating with them produces floating-point artifacts. Even opening the ply file in meshlab shows floating-point noise.

A better solution is to maintain the reconstruction in its local coordinate system and keep track of the transformation from that coordinate system to the one specified by the proj.4 line. We have added an export option that prints the matrix transformation from the original reconstruction coordinate system to the proj.4 one.

[dakotabenjamin]

I tested this on the aukerman test dataset but it doesn't work for EPSG:3734 Ohio State Plane. a The points ended up about 500km west of where they should be. It works fine for UTM17N. Could it be an issue with the ftUS attribute?

[dakotabenjamin]

So ODMGeoreferencing also produces a transformation matrix, and I'll note that the one for aukerman is slightly different than produced by your script. (both using UTM 17N)
odm_georeferencing:

[ 0.99837108,	0.00777421,	0.00097746,	0.54773735 ]
[ -0.00777667,	0.99836832,	0.00253651,	-0.54857154 ]
[ -0.00095768,	-0.00254405,	0.99839813,	0.55105865 ]
[ 0.00000000,	0.00000000,	0.00000000,	1.00000000 ]

extract_geocoords:

0.999611261999     0.00866169237997   0.0  437036.724336
-0.00866169203073  0.999611262232     0.0  4572789.22917
7.72997736931e-08  7.82310962677e-08  1.0  9.31322574615e-10
0.0                0.0                0.0  1.0

Basically the way odm_georef works (for exif) is that it extracts UTM coordinates by

1. converting each camera coordinate to UTM
2. finding an average point
3. saving the coordinate system, the average point, and each utm coordinate offset by the average.

then the actual georeferencing part finds the optimal camera triplet and calculates a transformation matrix from that, then applies the matrix to each point or mesh vertex.

[dakotabenjamin]

Did you manage to find a solution to the state plane issue?

[paulinus]

@dakotabenjamin what do you use for checking where the points end up?
I can convert the reconstruction to the lcc coordinates and get some numbers but I miss a tool to check if the numbers are correct.

[dakotabenjamin]

I just did a visual check on a known location using QGIS and a basemap layer. I did a rough measurement of the distance using a built-in calculator

[paulinus]

I guess that is loading the orthophoto in QGIS, not the point cloud. Need to try that. I was playing with the point cloud only.

Is ODM already using the geocoords_transformation.txt file in odm_georeferencing?

[dakotabenjamin]

Sorry, I wasn't clear. I actually loaded the points as a csv. I'm nearly done getting the geocoords to work with georeferencing, but it's not in master yet.

[dakotabenjamin]

@paulinus any update on this? I could also try my hand if you help point me in the right direction.