| The labelled pointclouds are released as pcd files. To read an pcd file into a numpy array, we recommend the package pypcd. |
|
|
|
|
| How to use evo to evaluate the pose I get with the pose_inW.csv? |
|
|
| It says clearly. it must have 8 entry. TUM expects: timestamp x y z q_x q_y q_z q_w |
|
|
| what you have is num t x y z qx qy qz qw |
|
|
| 1num 2t 3x 4y 5z 6qx 7qy 8qz 9qw |
|
|
| I think the answer is pretty clear. Delete First column |
|
|
|
|
|
|
|
|
| How to obtain annotated Lidar with corresponding pose files? |
|
|
| I have uploaded the poses of the pointclouds in each ground truth folder. Please check them and let me know if there is any issue. |
|
|
|
|
|
|
| Thank you very much for your work! |
| In many sequences the gps data seems broken (e.g. ntu_day_01). I was wondering if this is really the data you got or if simply something went wrong when splitting the bags. |
|
|
| Best regards, |
| Louis |
|
|
|
|
| Hi Louis. For GPS, there is no guarantee of availability. We tried to collect data, but most of the time, GPS suffers from multi-path issues or fails to lock. That's why we had to survey the large area and perform CEVA fitting of the trajectory for ground truth. If you take a look at the campus, you'll see that most places have tall structures. One suggestion is to fit the available GPS data with the LiDAR trajectory and then generate pseudo-GPS data. This approach should be more reliable. |
|
|
|
|
| Hi, |
|
|
| For the NTU and KTH sequences, the prior maps were aligned with GPS coordinates of the real locations. |
|
|
| The original prior maps have very large coordinates, presumably in some global position convention, which we have added some offsets to it to produce the ground truth |
|
|
| NTU: |
| geo_offset = [-10900.00, -36000.00, -20]; |
|
|
| KTH: |
| geo_offset = [-154100.0, -6581400.0, 0.0]; |
|
|
| I suppose you can add these to the ground truth coordinates and use some conversions to get back the lat-lon coordinates. But I am not sure what the formulas are. |
|
|
|
|
|
|
|
|
| Descrete time ground truth |
| For those who only need a traditional discrete-time ground truth for benchmarking SLAM methods, you can use the pose_inW.csv associated with each sequence. This data is sampled from the continuous time ground truth at 0.1s interval. The content looks like the following: |
|
|
| num t x y z qx qy qz qw |
| 11 1644823132.49211 49.2606317111444 107.371797989247 7.63580957239259 0.936118452267473 -0.351663294301812 0.003894594980225 -5.3806028052E-05 |
| 12 1644823132.59122 49.2617548510814 107.371820422962 7.63634221875222 0.936400655699795 -0.350853723839147 0.001952716665257 -0.007187723619002 |
| 13 1644823132.69085 49.2627442641658 107.370734597833 7.63708972884333 0.937130644015062 -0.348934531427190 0.005541781558079 -0.000370786684933 |
| 14 1644823132.79213 49.2634361009274 107.369692804349 7.63756375797164 0.935347736233029 -0.353687216878511 0.004396339541646 -0.003261462726211 |
| 15 1644823132.89130 49.2637252742628 107.368872120271 7.63750418262036 0.936261362661676 -0.351231145451445 0.005677919353789 -0.004370867600258 |
| ... |
| Here, num is the index of the lidar pointcloud in the bag file, t is the time stamp, and (x, y, z, qx, qy, qz, qw) is the pose of the body frame wrt to the world frame. |
|
|
|
|
| {0: 'barrier', |
| 1: 'bike', |
| 2: 'building', |
| 3: 'chair', |
| 4: 'cliff', |
| 5: 'container', |
| 6: 'curb', |
| 7: 'fence', |
| 8: 'hydrant', |
| 9: 'infosign', |
| 10: 'lanemarking', |
| 11: 'noise', |
| 12: 'other', |
| 13: 'parkinglot', |
| 14: 'pedestrian', |
| 15: 'pole', |
| 16: 'road', |
| 17: 'shelter', |
| 18: 'sidewalk', |
| 19: 'stairs', |
| 20: 'structure-other', |
| 21: 'traffic-cone', |
| 22: 'traffic-sign', |
| 23: 'trashbin', |
| 24: 'treetrunk', |
| 25: 'vegetation', |
| 26: 'vehicle-dynamic', |
| 27: 'vehicle-other', |
| 28: 'vehicle-static'} |
