Monocular scale error accumulation

[yada222]

Hello everyone, I would like to ask a question. When I try to create images with a monocular camera（whether it’s a pinhole camera or a fisheye camera）, there will always be a situation where the scale error becomes larger and larger as the turns increase. For example, the following is the test result of the public data set TUM_outdoor4. Although the tester is walking at a constant speed with the camera, the key frames and map points are all piled together。

My own cameras have also been calibrated and the calibration parameters have been modified。 I also tried some modifications, such as increasing the number of ORB points, filtering out orb points that are far away, etc., but the effect is not satisfactory. Has anyone encountered a similar problem? Do you know the reason or find a solution?

[CSU-NXY]

Same problem here. While test on KITTI sequence 01 with monocular mode, the scale suddenly changed.

[treememory]

same problem. in same dataset, after multi scale optimization, scale tends to 1. sometimes scale turn larger or samller. sometimes, Converge to a const scale to real world. anyone fix it? thank you~