You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
First of all thanks for having put in the effort for this useful dataset. It seems however that the latitude and longitude of the images center point are not correct. It seems that you have supplied the setpoints of the UAV trajectory rather than actual measured coordinates. When I plot out the trajectory, it follows a perfectly straight line, but when I zoom in on any specific image, the center point is off from the given coordinate. Probably, due to wind disturbances the drone will have flown slightly differently than the setpoint trajectory.
I have labelled a portion of the dataset manually, among which a portion of the 6th map trajectory. The manual label procedure was overlaying a cross on each image, and placing a marker in Google Maps at the center point (by eyeballing) and copying the coordinates. This picture shows the comparison between the manually labelled trajectory and the dataset trajectory. My algorithm performs better on the manually labelled ground truth:
Furthermore, for a few maps I have taken around 10% of images randomly, labelled them, and computed the error against the coordinates from the dataset. These are the error values I found for these maps:
Map 2: 24.16m
Map 4: 47.73m
Map 10: 30.76m
Map 11: 25.82m
The text was updated successfully, but these errors were encountered:
First of all thanks for having put in the effort for this useful dataset. It seems however that the latitude and longitude of the images center point are not correct. It seems that you have supplied the setpoints of the UAV trajectory rather than actual measured coordinates. When I plot out the trajectory, it follows a perfectly straight line, but when I zoom in on any specific image, the center point is off from the given coordinate. Probably, due to wind disturbances the drone will have flown slightly differently than the setpoint trajectory.
I have labelled a portion of the dataset manually, among which a portion of the 6th map trajectory. The manual label procedure was overlaying a cross on each image, and placing a marker in Google Maps at the center point (by eyeballing) and copying the coordinates. This picture shows the comparison between the manually labelled trajectory and the dataset trajectory. My algorithm performs better on the manually labelled ground truth:
![gt_traj_comparison](https://private-user-images.githubusercontent.com/35661626/391475025-07a9153d-57e0-437c-aa61-9ab46c04d829.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.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.1jNKqH9nodIA4Uou21OYaT9Td1G7ZFm8_Khy4uCqMw4)
Furthermore, for a few maps I have taken around 10% of images randomly, labelled them, and computed the error against the coordinates from the dataset. These are the error values I found for these maps:
Map 2: 24.16m
Map 4: 47.73m
Map 10: 30.76m
Map 11: 25.82m
The text was updated successfully, but these errors were encountered: