Skip to content

Latest commit

 

History

History
75 lines (65 loc) · 2.52 KB

README.md

File metadata and controls

75 lines (65 loc) · 2.52 KB
Raw

dependency

  • numpy
  • scipy
  • opencv-python
  • tqdm
  • scikit-image
  • pykdtree

Usage

Get groundtruth depthmap (skip this step if the depthmaps are provided)

python ptc2depthmap.py --output_path <output_path> \
    --input_path <input_pointcloud_path> \
    --calib_path <KITTI_calib_folder> \
    --image_path <KITTI_image_folder> \
    --split_file --split_file <split_file>  --threads <thread_number>

Run batch GDC on predicted depth maps

python main_batch.py --input_path <path_to_pred_depth_map> \
    --calib_path <KITTI_calib_folder> \
    --gt_depthmap_path <path_to_gt_depth_map>\
    --output_path <output_path> --threads <thread_number> \
    --split_file <split_file> (<otherargs>)

Get pointclouds from corrected depth map

python depthmap2ptc.py --output_path <output_path> \
    --input_path <input_depthmap_path> \
    --calib_path <KITTI_calib_folder> \
    --split_file <split_file> --threads <thread_number>

The split file contains ids of scenes on which we are to run GDC, e.g.

000000
000003
000007
000009
000010
000011
...

Sparcify point clouds

Since PointRCNN model exploits the sparcity of the point-clouds, when we apply it on pseudo-LiDAR point-clouds (which are dense), we follow the procedure below to slice the point-clouds to make it have similar sparse property of LiDAR. We found using the sparsed pseudo-LiDAR point-clouds improves the 3D detection performance of PointRCNN model.

Sparse pseudo-LiDAR point-clouds to 64 lines

python sparsify.py --calib_path <KITTI_calib_folder> \
    --image_path <KITTI_image_folder> --ptc_path <pointcloud_folder> \
    --split_file <split_file> --W 1024 --slice 1 --H 64 --threads <thread_number>

Simulate 4-beam LiDAR

To extract 4-line LiDAR from the velodyne data provided by KITTI, run

python sparsify.py --calib_path <KITTI_calib_folder> \
    --image_path <KITTI_image_folder> --ptc_path <pointcloud_folder> \
    --W 1024 --H 64 --line_spec 5 7 9 11 \
    --store_line_map_dir <temporary_store_path>

Note that this sparcify method is slightly diffrent from sparsify.py in the src folder, in that we store the 4 beam LiDAR angular map and put it back after GDC. The <temporary_store_path> is used to store the 4-beam ground-truth in angular map.

Sparsify the corrected point clouds

python sparsify.py --calib_path <KITTI_calib_folder> \
    --image_path <KITTI_image_folder> --ptc_path <corrected_pointcloud_folder>\
    --W 1024 --slice 1 --H 64 --fill_in_map_dir <temporary_store_path>\
    --fill_in_spec 5 7 9 11