To enrich the dataset in the moving object segmentation (MOS) task and to reduce the gap of different data distributions between the validation and test sets in the existing SemanticKITTI-MOS dataset, we automatically annotated and manually corrected the KITTI-Road dataset.
More specifically, we first use auto-mos labeling method (link) to automatically generate the MOS labels for KITTI-Road data. We then use a point labeler (link) to manually refined the labels.
We follow semantic SLAM SuMa++ to rename the sequences of KITTI-Road data as follows.
raw_id -> seq_id
2011_09_26_drive_0015 -> 30
2011_09_26_drive_0027 -> 31
2011_09_26_drive_0028 -> 32
2011_09_26_drive_0029 -> 33
2011_09_26_drive_0032 -> 34
2011_09_26_drive_0052 -> 35
2011_09_26_drive_0070 -> 36
2011_09_26_drive_0101 -> 37
2011_09_29_drive_0004 -> 38
2011_09_30_drive_0016 -> 39
2011_10_03_drive_0042 -> 40
2011_10_03_drive_0047 -> 41
We provide a simple download and conversion script utils/download_kitti_road.sh, please modify the DATA_ROOT path and manually move the result folder sequences to the target folder.
And you need to download the KITTI-Road-MOS label data annotated by us, the pose and calib files from here (6.4 MB) [Remap the label to 9 and 251, consistent with the SemanticKITTI-MOS benchmark]. old version here (6.1 MB).
We organize our proposed KITTI-Road-MOS using the same setup and data structure used in SemanticKITTI-MOS:
DATAROOT
├── sequences
│ └── 30
│ ├── calib.txt # calibration file provided by KITTI
│ ├── poses.txt # ground truth poses file provided by KITTI
│ ├── velodyne # velodyne 64 LiDAR scans provided by KITTI
│ │ ├── 000000.bin
│ │ ├── 000001.bin
│ │ └── ...
│ ├── labels # ground truth labels from us
│ │ ├── 000000.label
│ │ ├── 000001.label
│ │ └── ...
│ └── residual_images_1 # the proposed residual images
│ ├── 000000.npy
│ ├── 000001.npy
│ └── ...