Running the TUM-VI dataset in stereo mode has large angular errors during evaluation

[successkaisadadi]

When I used run_ros_tumvi.sh to run the TUM-VI dataset, I found that the angle error reached 3-5deg. In other papers citing OpenVINS, the angle error is less than 2deg. Is there a similar solution?

[goldbattle]

I have not done an evaluation in a while. Will try to take a look as soon as I get time, thanks for opening an issue.

[successkaisadadi]

Thanks for your reply, do you have any suggestions for now or what I need to try to solve this problem?

[goldbattle]

What OpenCV version are you using?

[goldbattle]

It looks like this is an issue with accidently deleting IMU readings. Not sure how why I used the original logic of the oldest feature take. If there is a drop in tracking (as in the beginning of the TUM datasets), all IMU readings will be deleted / dropped. I will push a fix ASAP but you can change these lines:

open_vins/ov_msckf/src/core/VioManager.h

Lines 80 to 85 in bf0ada9

	// Get the oldest camera timestamp that we can remove IMU measurements before
	// Then push back to our propagator and pass the IMU time we can delete up to
	double oldest_time = trackFEATS->get_feature_database()->get_oldest_timestamp();
	if (oldest_time != -1) {
	oldest_time += params.calib_camimu_dt;
	}

To the following:

  // The oldest time we need IMU with is the last clone
  // We shouldn't really need the whole window, but if we go backwards in time we will
  double oldest_time = state->margtimestep();
  if (oldest_time > state->_timestamp) {
    oldest_time = -1;
  }

Let me know if this fixes the issue for you.

[goldbattle]

I am not sure, as I have not looked into the problem and there are not many details. I guess one large change is that we have added dynamic initialization in v2.6, which in the past wouldn't have been used.

…

On Sun, May 8, 2022 at 3:56 AM gaoyunqi ***@***.***> wrote: Thanks for your reply, do you have any suggestions for now or what I need to try to solve this problem? — Reply to this email directly, view it on GitHub <#246 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAQ6TYVJUJ2SPVZVH2MCK2TVI5XT3ANCNFSM5VKQZOAQ> . You are receiving this because you commented.Message ID: ***@***.***>