Estimated trajectories by milliego has huge drifts.

[JzHuai0108]

Dear authors,

Thank you for sharing the code, milliego model, and dataset. We are running the milliego handheld model on the handheld sequences. The resulting trajectories are heartbreaking. For 2020-01-28-11-10-12, the resulting trajectory with gt is like below For 2020-01-28-11-15-11, the resulting trajectory is like below,

We download the cnn model from here. The handheld milliego model is from here. The handheld sequences are downloaded from here. We create the h5 files for the above two sequences with the scripts from here.

python extract_files.py
python process_radar_handheld_uav.py
python os_create_dataset_1_slaves.py

The master gap we put in the config.yaml is 4 which is the default value. Then we create the docker container as instructed here. We set the IMU_LENGTH=20 at here which is the default value used in the config.yaml. Other values like 25 will cause an error complaining the incompatible shape of numpy arrays, 3200 vs 2560. Then we get the above results by running test.py. We also try to generate h5 files with master_gap = 5, but also setting IMU_LENGTH=20 as other values say 25 cause the imcompatible size error. The results are much worse, the trajectory drifts many order of magnitude away in position.

The odombeyondvision paper reports impressive results for milliego unlike what we have got. I think we must have missed some key points. Can you please pinpoint the cause for us? Our code for generating the above results is at here on branch main.

Regards,

[JzHuai0108]

Please refer to here. Try the milliEgo model '18' over there.