MOE-Dataset: A Dense LiDAR Moving Event Dataset

Project Page | [Paper] | Competition | Dataset

Dataset

Sequences Format

We collect multiple sequences from simulators/real urban campus containing both high-density moving objects for indoor/outdoor scenes. An overview for these sequences are as follows. (Actually there are much more moving/non-moving points than you can in these animations since we lower the opacity of pointcloud to make these visualization more human-friendly visible.)

00	01	02	03	04

05	06	07	08	09

The folder of whole dataset is like this:

MOE-Dataset
├── 00
│   ├── gt_poses.txt
│   ├── label/
│   └── pcd/
├── 01
│   ├── gt_poses.txt
│   ├── label/
│   └── pcd/
├── 02
│   ├── gt_poses.txt
│   ├── label/
│   └── pcd/
├── 03
│   ├── gt_poses.txt
│   ├── label/
│   └── pcd/
... ...
├── 08
│   ├── gt_poses.txt
│   ├── pcd/
|
├── 09
│   ├── gt_poses.txt
│   ├── pcd/
|___ ___

Pointcloud

Each scan frame is provided as xxxxxx.pcd with x,y,z information. We choose to privode the raw *.pcd file for eaiser visualization and loading, although binary format adopted by SemanticKITTI may be more storage efficient. It's very easy to read the pointcloud files using library like Open3D.

Label

For the label file xxxxxx.txt, each row is the hierarchical motion label for each point in the cooresponding pcd file, which is defined as:

Thus, each row in the label file xxxxxx.txt is

Moveable_ID Moving_Status Class_ID

We denote the moveable point as 1 otherwise 0, the moving status as 1 otherwise 0. BTW, we only provide the semantic class ID for moving points otherise -1 for non-moving points.

Ground truth Pose

For each sequence, the ground truth poses for all scans are store in gt_poses.txt. Each row in gt_poses.txt represents the pose for the cooresponding ordered scan. For example, first row represents the pose for 000000.pcd in that sequence. We adopt the KITTI format to reprensent the pose, that is the first 3 rows of elements in the 4x4 transformation matrix.

# Each row in gt_poses.txt
R11 R12 R13 x R21 R22 R23 y R31 R32 R33 z

# The cooresponding 4x4 transformation matrix
T = 
[
    R11 R12 R13 x
    R21 R22 R23 y
    R31 R32 R33 z
    0   0   0   1
]

Download

We put the zipped sequence file in the release of this repository https://github.com/DeepDuke/MOE-Dataset/releases/tag/publish_dataset . Please download and unzip them. Note: Some pcd file of sequences may be splitted into several zipped file due to the single file limitation of Github is 2GB. Please merge the unzipped files into same sequence folder.

Benchmark

We evaluate some SOTA algorithms on sequences 00,01,02 to set up a reference benchmark for their performance. We test 3 offline non-learning methods - Removert, ERASOR, and Octomap; 3 online non-learning methods - Dynablox, DOD, M-detector; two representives for different branches of learning based methods - MotionBEV and InsMOS. For non-learning based methods, we tried to tune those params on each sequence. For learn-based methods, we test them using model trained on the larger scale dataset SemanticKITTI to validate their generalization ability. Actually current learning-based methods are verry easy to overfit on single structure dataset (like urban high ways), and not so robust on different structured high-density scenes in this dataset.

ERASOR on Seq 00	MotionBEV on Seq 01	InsMOS on Seq 02

We compute the mean IoU metric over sequences {00, 01, 02} for these algorithms. We refer the moving point as positve while non-moving point as negative.

IoU = TP / (TP + FP + FN)

Seq	Removert	ERASOR	Octomap	Dynablox	DOD	M-detector	MotionBEV	InsMOS
00	0.297	0.378	0.328	0.320	0.786	0.305	0.002	0.495
01	0.028	0.028	0.031	0.195	0.142	0.174	0.055	0.282
02	0.421	0.627	0.652	0.492	0.595	0.044	0.069	0.379
mIoU	0.249	0.344	0.337	0.336	0.508	0.174	0.042	0.385

Competition and Leaderboard

To facilitate the related community, we utilize part sequences {05,06,07,08,09} to host a Moving Event Detection with LiDAR on CodaLab. So our dataset will only provide motion label for sequences {00, 01, 02, 03, 04}, for sequences {05, 06, 07, 08, 09} we will only publish the pointcloud files and ground truth poses.

The link for our competition is at https://codalab.lisn.upsaclay.fr/competitions/18028. If you are interested in it, please feel free to take part in it!

Issues

If you have any question about this dataset, please feel free to open an issue for discussion. TKS!

Acknowledgement

We want thank the following open source code and dataset.

• Removert
• ERASOR
• An improved version of Octomap
• Dynablox
• DOD (dynamic_object_detection)
• DynamicMap_Benchmark
• M-detector
• MotionBEV
• InsMOS
• SemanticKITI Dataset

Citation

Citation format is comming soon!

If you find this dataset useful to you, please feel free to give us a star! TKS!

Star History