triangulation with know poses

[Shine233]

Hi

Thanks for the gread work!
Since i already have prior poses with real scale that lack of accuracy in autodriving scene, I prefer to appIy triagulation with those prior poses and then run ba to obatain more accurate extrinsic paramters with real scale.
I have noticed that there are several ways to run with prior known poses, such as using GPS prior in image EXIF, or using openMVG_main_SfMInit_ImageListingFromKnownPoses to load prior poses. However, using the first method, the prior pose only constrian the translation which will cause a random angle error compared with INS (mainly in roll in vehicle coordinate). Besides, i tried to use openMVG_main_SfMInit_ImageListingFromKnownPoses with ETH_3D format which poses are camera2world, if i use openMVG_main_ComputeStructureFromKnownPoses func, the BA process run only 1 iter that is not enough; if i use openMVG_main_SfM INCREMENTAL func, the scale seems to be eliminated.

I would be really appreciate it if you could give some advise on it.
Many thanks

[pmoulon]

YOu can run openMVG_main_SfMInit_ImageListingFromKnownPoses and then reun SequentialSfM 2 with the mode known poses.

So run main_SfM with INCREMENTALV2 and EXISTING_POSE https://github.com/openMVG/openMVG/blob/develop/src/software/SfM/main_SfM.cpp#L286

[Shine233]

Thanks for your explaination!
As you mentioned above, i have tried your method and other 2 methods, however they all looks bad.
Here is my scene, a city view road with a straight trajectory captured by a vehicle, about 200 meters with 129 images contained.
Next is my experiments:

1. Initialized with openMVG_main_SfMInit_ImageListingFromKnownPoses and then use SequentialSfM 2 with -S .
2. The same with exp1, but add -P option in openMVG_main_SfM INCREMENTALV2. I have modifed the main_SfMInit_ImageListing so that sfm_data.json contains prior pose in views as well as extrinsics.
3. Use normal openMVG_main_SfM INCREMENTAL with -P option to recover the real scale.
4. Use normal openMVG_main_SfM GLOBAL with -P option to recover the real scale.

However, the evaluation of these 4 expriments have something in common. The errors compared with GT poses is large at the begin and end of the trajectory, and errors decrease at the middle section.
Hence i have several questions,

1. For a small forward moving scene(less than 500 meters and less than 300 images), which method do u recommoned to fix the above problem?
2. Does prior poses in views really works if I already use the same prior poses to initialized the sfm_scene with -S EXISTING_POSE in openMVG_main_SfM INCREMENTALV2?
3. Since the above scene is a straight forward-moving case, the sim3 residuals seem lack of constrains in rotation，will this be a reason to cause bad result?
4. If for a quit large scene, do u have any suggestion to do sfm? What about to use distributed sfm?

Sorry for the late reply. Hope you can give me some advices.

Many thanks!