Time cost comparison
alternating_converge parameter.
By lowering this parameter, alternating optimization can be completed faster, but convergence to 10% rate is already fast, and the optimization can sometimes take a long time due to an inaccurate optimization.
Anyway, lowering the parameter reduces the mean time.
alternating_converge parameter.
Time comparison in desktop (AMD Ryzen 9 3900x)
You can check the accuracy results in in parameter page.
Time comparison in NVIDIA Xavier
DynaVINS(dev) is DynaVINS stereo-IMU mode with 'margin_feature_thresh' parameter.
Current verson on Github is 'dev' version.
Using alternating_converge parameter (0.1 for those experiments), DynaVINS(dev) shows a very fast speed compared to the initial version, and sometimes even faster than the existing VINS-Fusion.
Time consumption comparison in the temporal static sequence of our dataset.
Time consumption comparison in the E-shape sequence of our dataset.
Because our loop closure module optimizes the trajectory after grouping the keyframes and clustering the matching results, even though a single optimization takes a long time, the cumulative time is similar to or less than that of VINS-Fusion.
VINS-Fusion combined with Switchable Constraints performs the same number of optimizations as VINS-Fusion. But it consumes more time for each optimization, so it requires more time than others.