search

Hilti-Research / hilti-slam-challenge-2022

51 stars 1 forks source link

Ground-truth localization details #11

Closed tpet closed 1 year ago

tpet commented 2 years ago

Thanks for this interesting dataset. Could you provide details on how was the ground-truth localization obtained? Was is similar to the previous year challenge (measured by total station)?

hilti-challenge commented 2 years ago

This year we used a survey-grade laserscanner (Z+F Imager 5016) to scan all locations. The individual groundtruth points are extracted from the scan (using laserscanner targets)

tpet commented 2 years ago

And how are you going to obtain ground-truth localization from that?

hilti-challenge commented 2 years ago

similar to last year, but we replaced the total station with laser-scanner targets like the one in the picture. image

We "revisited" the same physical point on the ground with our lidar-camera-hardware

YoshuaNava commented 2 years ago

Was this the source of the ground truth trajectory for exp03-06 (which is said to be different and have an error of 1-3cm in the website), or for the whole challenge (to align against before computing the error at the control points)?

IamPhytan commented 2 years ago

How did you managed to scan targets that are not visible from the start of the runs (like in the cupola) ? Did you compute the transformation between each pose of the scanner ? If yes, how ?

bedaberner commented 2 years ago

Was this the source of the ground truth trajectory for exp03-06 (which is said to be different and have an error of 1-3cm in the website), or for the whole challenge (to align against before computing the error at the control points)?

For the ground truth trajectory IPC was used to compare the lidar data to the Laserscan. This allowed us to create a dense trajectory but includes one of the sensors into the ground truth. When comparing the dense ground truth generated by this method to the static targets measured directly in the laserscan, errors between 1-3cm resulted. This method was not at all involved in the final evaluation.

How did you managed to scan targets that are not visible from the start of the runs (like in the cupola) ? Did you compute the transformation between each pose of the scanner ? If yes, how ?

We used many scans to create a complete pointcloud of the areas. They were combined using the usual workflow of the laserscanner (Target, ICP and plane matching, search for scantra if you are interested in the actual software used).

You can also watch this video to see a bit of the scanning and the resulting laserscan.