Closed mateosss closed 10 months ago
I don't have any, but the script you can use to generate them for your hardware are in the repo: ov_msckf/scripts/
After a while, I got to compare OpenVINS against Basalt on the EuRoC datasets using the xrtslam-metrics tooling. I used the default euroc settings, and ran it without ROS. Here are the results for anyone that might find it useful (you can cite the Monado SLAM datasets if needed).
Note that I removed the EMH04 dataset from the average row since OpenVINS diverged on it.
Average completion percentage [%]
(Percentage of the dataset for which the system gives a trajectory estimate to evaluate)
Basalt | OpenVINS | |
---|---|---|
EMH01 | 100% | 75.60% |
EMH02 | 100% | 74.30% |
EMH03 | 100% | 85.59% |
EMH05 | 100% | 81.34% |
EV101 | 100% | 96.15% |
EV102 | 100% | 94.21% |
EV103 | 100% | 93.30% |
EV201 | 100% | 95.66% |
EV202 | 100% | 96.51% |
EV203 | 100% | 92.89% |
[AVG] | 100% | 88.55% |
Absolute trajectory error (ATE) [m]
Basalt | OpenVINS | |
---|---|---|
EMH01 | 0.076 ± 0.021 | 0.070 ± 0.037 |
EMH02 | 0.048 ± 0.024 | 0.111 ± 0.044 |
EMH03 | 0.060 ± 0.022 | 0.125 ± 0.054 |
EMH05 | 0.145 ± 0.043 | 0.200 ± 0.089 |
EV101 | 0.045 ± 0.018 | 0.061 ± 0.021 |
EV102 | 0.054 ± 0.017 | 0.080 ± 0.038 |
EV103 | 0.056 ± 0.014 | 0.059 ± 0.026 |
EV201 | 0.036 ± 0.013 | 0.048 ± 0.026 |
EV202 | 0.040 ± 0.014 | 0.042 ± 0.018 |
EV203 | 0.156 ± 0.059 | 0.117 ± 0.035 * |
[AVG] | 0.072 ± 0.024 | 0.091 ± 0.039 |
Relative trajectory error (RTE) [m]
Basalt | OpenVINS | |
---|---|---|
EMH01 | 0.005118 ± 0.002989 | 0.011286 ± 0.006887 |
EMH02 | 0.004231 ± 0.002249 | 0.009638 ± 0.004916 |
EMH03 | 0.011682 ± 0.007597 | 0.020609 ± 0.011800 |
EMH05 | 0.010716 ± 0.006293 | 0.020645 ± 0.012218 |
EV101 | 0.013046 ± 0.006272 | 0.014264 ± 0.006360 |
EV102 | 0.011127 ± 0.004774 | 0.015966 ± 0.009288 |
EV103 | 0.012287 ± 0.006093 | 0.014620 ± 0.007033 |
EV201 | 0.003423 ± 0.001966 | 0.008418 ± 0.006396 |
EV202 | 0.007064 ± 0.003854 | 0.009062 ± 0.004893 |
EV203 | 0.026494 ± 0.021414 | 0.024541 ± 0.016509 * |
[AVG] | 0.010519 ± 0.006350 | 0.014363 ± 0.008011 |
*: The original EV203 dataset has missing frames, it was modified to only use valid stereo frames for OpenVINS to improve scores.
An interesting example of trajectories with the two systems on EV101, notice how the loop closures cause abrupt corrections.
Hello, are there any up to date reports of the system performance on euroc and tumvi datasets? I wasn't able to find such numbers, only the ones from the original 2020 paper which I imagine might have changed.
Thank you.