You may need OpenCV, pcl, Eigen3 installed before compiling.
mkdir build
cd build
cmake ..
make -j8
cd ..
These binaries select keyframes based on the interval distance between two consecutive frames.
You may specify the interval manually, e.g. 1 or 2, in meters.
After running the binaries, you will have files organized as follows:
[keyframes_root_dir]
├ ...
└ selected_keyframes_xxm/
├ keyframe_point_cloud/ <- folder for keyframe point clouds in pcd format
├ keyframe_pose.csv <- gt poses for all selected keyframes
└ keyframe_pose_format.csv <- format for one entry in the gt poses
Extract keyframes from MulRan dataset using same spatial distance for consecutive keyframes.
Usage:
./build/mulran_point_cloud_select [dataset_root_dir] [keyframe_dist_interval](default=2)
[dataset_root_dir] should be organized as follows:
[dataset_root_dir]
├ sensor_data
│ ├ Ouster/
│ └ ouster_front_stamp.csv
└ global_pose.csv
Extract keyframes from KITTI Odometry dataset using same spatial distance for consecutive keyframes.
Usage:
./build/kitti_point_cloud_select [dataset_root_dir] [keyframe_dist_interval](default=2)
[dataset_root_dir] should be organized as follows:
[dataset_root_dir]
├ velodyne/
├ times.txt
└ global_pose.txt
Extract keyframes from Oxford Radar RobotCar dataset using same spatial distance for consecutive keyframes.
Usage:
./build/oxford_point_cloud_select [dataset_root_dir] [keyframe_dist_interval](default=2)
[dataset_root_dir] should be organized as follows:
[dataset_root_dir]
├ velodyne_left/
├ velodyne_left.timestamps
└ gps/
└ ins.csv
Apply a regid transform on a single input pcd pointcloud, and generate BEV image for both original and transformed point clouds.
Usage:
# translation (trans_x, trans_y, trans_z) in metre
# rotation on yaw axis, in degree
./build/cloud_manip [input_pcd] [trans_x] [trans_y] [trans_z] [theta]
Generate BEV image all point clouds within the specified keyframes folder.
Usage:
./build/batch_cloud_manip [keyframes_root_dir]
This binary generates ground-removed point clouds, single & multi layer BEV images and creates geometric distance-based labels for each point cloud.
After running the binary, you will have files organized as follows:
[keyframes_root_dir]
├ ...
├ non_ground_point_cloud/ <- folder for ground-removes point clouds in pcd format
├ output_multi_bev/ <- folder for multi-layer BEV images
└ output_single_bev/ <- folder for single-layer BEV images
Single-layer BEV images are provided in csv and png format, and multi-layer BEV images are provided in raw binaries bin and png format.
Usage:
./build/batch_multi_bev_gen [keyframes_root_dir] [sensor_type]
[keyframes_root_dir] should be organized as follows:
[keyframes_root_dir]
├ keyframe_point_cloud/ <- folder for selected point clouds in pcd format for each frame
├ keyframe_pose.csv <- 6-DoF pose for each frame
└ keyframe_pose_format.csv <- 6-DoF pose format description
[sensor_type] could be HDL_32E, HDL_64E or OS1_64.
Two-stage pose estimation using flattened 2D point clouds for a single matching pair.
Usage:
# two point clouds to be aligned: query_pcd and match_pcd.
# initial guess for relative rotation on yaw axis, in degree.
./build/top_part_registration [query_pcd] [match_pcd] [yaw_initial_guess]
Perform two-stage pose estimation on a batch of point cloud pairs.
Usage:
# need a match result text file "match_result.txt" descripting all matching pairs.
# each entry contains one matching pair: "query_idx match_idx yaw_initial_guess"
./build/batch_top_part_registration [match_result_text_file] [point_cloud_dir]
Use Keyframes Extractors to extract keyframes from your datasets. 3 different extractors are provided to handle different datasets.
Use BEV Generator to convert the pcd-format keyframes to BEV images. The csv-format single-layer BEV images and their corresponding ground truth poses will be used in FreSCo for place retrieval and performance estimation.