Hello.
We have a forward-motion dataset taken at 50 Hz.
We were able to run LSD-SLAM to the end of the dataset, but the resulting camera trajectory shows quiet shorter length and less turning angles than the ground-truth path.
We think the issue is caused by scale. The keyframe scale (camToWorld_estimate.scale() in KeyFrameGraph:addKeyFrame) gradually becomes smaller value (from scale of 1 to 0.01) from the beginning of path to the end.
Our dataset does not have the global loop-closing, so the scale drift and error could become large.
But I found this post (https://github.com/tum-vision/lsd_slam/issues/101). According to @JakobEngel, one reason that the KITTI dataset does not work well in LSD-SLAM is due to the dataset is forward-motion dataset.
I would like to ask whether monocular LSD-SLAM does not generally work well with forward-motion dataset, and whether the forward-motion dataset could cause scale error (i.e., gradually becoming smaller scale).
Hello. We have a forward-motion dataset taken at 50 Hz. We were able to run LSD-SLAM to the end of the dataset, but the resulting camera trajectory shows quiet shorter length and less turning angles than the ground-truth path.
We think the issue is caused by scale. The keyframe scale (camToWorld_estimate.scale() in KeyFrameGraph:addKeyFrame) gradually becomes smaller value (from scale of 1 to 0.01) from the beginning of path to the end.
Our dataset does not have the global loop-closing, so the scale drift and error could become large. But I found this post (https://github.com/tum-vision/lsd_slam/issues/101). According to @JakobEngel, one reason that the KITTI dataset does not work well in LSD-SLAM is due to the dataset is forward-motion dataset.
I would like to ask whether monocular LSD-SLAM does not generally work well with forward-motion dataset, and whether the forward-motion dataset could cause scale error (i.e., gradually becoming smaller scale).
Thank you!