jac99
commented
3 years ago Hi, our method works on pre-processed RobotCar datasets, prepared by authors of PointNetVLAD paper (available at: https://github.com/mikacuy/pointnetvlad). These point clouds are constructed by accumulating multiple scans from 2D LiDAR, accumulated over 20 meter drive. Such point clouds are downsampled to 4096 points, normalized so coordinates are in -1..1 range and ground plane points are removed. Which raw LiDAR data do you want to process? Original RobotCar dataset contains both 2D and 3D LiDARs.
- The individual scans from 2D LiDAR are rather small - they contain only about 500 points on one 2D plane. They are not informative enough for successful place recognition (or for other tasks).
- 3D LiDAR scans in original RobotCar dataset doesn't look good for place recognition.
If you want to use raw point clouds from a LiDAR for place recognition or similar purposes, you should look into Oxford Radar RobotCar (https://oxford-robotics-institute.github.io/radar-robotcar-dataset/). It contains LiDAR scans from Velodyne HDL-32E LIDAR. They are relative large point clouds with few thousand points covering area with app. 80 meter diameter. You will not need any special pre-processing for such point clouds. My network (and underlying Minkowski Engine) operates on such 3D points clouds. You can read the point cloud as triplets of (X, Y, Z) coordinates, quantize the coordinates and pass them through the network. You'll only need to change quantization step (default setting in my code is to use very small quantization step=0.01 - as point coordinates were normalized in -1..1 range). Raw scans have point coords within -80..80 range (or similar) and you need to change quantization step accordingly (0.3 meter looks reasonably good). To get reasonable results you'll need to re-train the model (as it was trained on accumulated point clouds from 2D LIDAR, which have different characteristics. Also network architecture will need tuning/amending to get the best results.
QLuanWilliamed
Hi, I am very interested in your nice work. I would like to how to process the raw dataset download from the Oxford RebotCar dataset that has a different date compared with your dataset. If you know how to pre-process the dataset, I wish you can give me some advice. Thanks a lot!