Feature extraction and projections

[msoiron]

Hi,
as I understand the orb feature extractor is applied to each frame without any preprocessing.

As we are using equirectangular frames, is it possible that the distortion in the top and bottom part of the video would lead to problems during feature detection in those parts and is that maybe the reason why they are masked in the provided yaml settings?

Would it be possible to unproject the whole frame and do the feature extraction within the spherical coordinate space (hence without distortions) and if yes, would that increase performance?

[ymd-stella]

As we are using equirectangular frames, is it possible that the distortion in the top and bottom part of the video would lead to problems during feature detection in those parts and is that maybe the reason why they are masked in the provided yaml settings?

Yes.

Would it be possible to unproject the whole frame and do the feature extraction within the spherical coordinate space (hence without distortions)

If needed, users can implement extra camera models (e.g. dual fisheye, catadioptric) with ease.

and if yes, would that increase performance?

I don't think the performance will be much different. If you are unsatisfied with the performance, there are other things you should be doing more of.

[chongyaaa]

Reducing baseline_dist_thr_ratio, will reduce the baseline?

[ymd-stella]

See

stella_vslam/src/stella_vslam/mapping_module.cc

Line 278 in f94dbda

// if the scene scale is much smaller than the baseline, abort the triangulation

[chongyaaa]

if (baseline_dist < baseline_dist_thr_ratio_ * median_depth_in_ngh) {
continue;
}
Reduce baseline_dist_thr_ratio_,will increase landmarks.

[chongyaaa]

Hello, how about the effect of your experiment?

[msoiron]

Hi,
I tried all those parameters but I could not see that it makes a large difference. I think the drift we observed is most likely due to poor lighting in one region of our video. So there is most likely not much we can do about it.