Table 2.

Available point cloud datasets for classification, segmentation, and object detection.

Datasets Name	Descriptions	Application Tasks
ModelNet40 [18]	It consists of 12,311 CAD models in 40 object classes.	3D object classification [48,50,51,79] and shape classification [45]
ShapeNet part [80]	There are 16,881 shapes represented by 3D CAD models in 16 categories with a total of 50 parts annotated.	Part segmentation [44,48,49,56,80], shapes generation [46], and representation learning [52]
Stanford 3D semantic parsing [81]	This dataset has 271 rooms in six areas captured by 3D Matterport scanners captured by Matterport Camera.	Semantic Segmentation [8,43,44,46,48,49,78,82]
SHREC15 [18]	There are 1200 shapes in 50 categories by scanning real human participants and using 3D modeling software [79]. Each class has 24 shapes and most of these shapes are organic with different postures.	Non-rigid shape classification [43]
SHREC16 [18]	It contains about 51,300 3D models in 55 categories.	3D shape retrieval [8]
ScanNet [83]	There are 1513 scanned and reconstructed indoor scenes.	Virtual scan generation [43], segmentation [48], and classification [48]
S3DIS [81]	It consists of 271 rooms in six areas captured by 3D Matterport scanners.	3D semantic segmentation [44,48] and representation
TU-Berlin [84]	It has sketches in 250 categories. Each category has 80 sketches.	Classification [48]
ShapeNetCore [85]	It has 51,300 3D shapes in 55 categories, which is indicated by triangular meshes. The dataset is labeled manually and a subset of the ShapeNet dataset.	3D shape retrieval task [78], 3D shape retrieval task [8], and classification [8]
ModelNet10 [18]	The 10-class of Model-Net (ModelNet10) benchmarks are used for 3D shape classification. They contain 4,899 and 12,311 models respectively.	Object classification [8] Shape classification [78]
RueMonge2014 [86]	The images are multi-view in high-resolution images from a street in Paris and the number of these images is 428.	3D point cloud labeling [56]
3DMatch Benchmark [87]	It contains a total of 62 scenes.	Point Cloud representation [54]
KITTI-3D Object Detection [88,89]	There are 16 classes, including 40,000 objects in 12,000 images captured by a Velodyne laser scanner.	3D object detection [20,23,24,90]
vKITTI [91]	This dataset includes a sparse point cloud captured by LiDAR without color information. It can be used for generalization verification, but it cannot be used for supervised training.	Semantic segmentation [51]
3DRMS [92]	This dataset comes from the challenge of combining 3D and semantic information in complex scenarios and was captured by a robot that drove through a semantically rich garden with beautiful geometric details.	Semantic segmentation [51]
Cornell RGBD Dataset	It has 52 labeled point cloud indoor scenes including 24 office scenes and 28 family scenarios with the Microsoft Kinect sensor. The data set consists of approximately 550 views with 2495 segments marked with 27 object classes.	Segmentation [14]
VMR-Oakland dataset	It contains point clouds captured by mobile platforms with Navlab11 around the Carnegie Mellon University (CMU) campus.	Segmentation [14]
Robot 3D Scanning Repository	The 3D point clouds acquired by Cyberware 3030 MS are provided for both indoor and outdoor environments. Heat and color information is included in some datasets.	Segmentation [14]
ATG4D [89]	There are over 1.2 million, 5,969, and 11,969 frames in the training, validation, and test datasets, respectively. This dataset is captured by a PrimeSense sensor.	Point object detection [20]
Paris-Lille-3D [60]	There are 50 classes in 143.1M point clouds acquired by Mobile Laser Scanning.	Segmentation and classification [60]
Semantic3D [93]	There are eight classes in 1660M point clouds acquired by static LIDAR scanners.	Semantic segmentation [93]
Paris-rueMadame [94]	There are 17 classes in 20M point clouds acquired by static LIDAR.	Segmentation, classification, and detection [94]
IQmulus [61]	There are 22 classes in 12M point clouds acquired by static LIDAR.	Classification and detection [61]
MLS 1 - TUM City Campus [95,96]	There are more than 16,000 scans captured by mobile laser scanning (MLS) in this dataset.	3D detection [95,96], city modeling [95,96], and 3D change detection