Page 19 - Read Online
P. 19
Wu et al. Intell Robot 2022;2(2):10529 I http://dx.doi.org/10.20517/ir.2021.20 Page 117
Figure 3. Typical networks for two categories of LiDAR-based tracker: (a) LiDAR-only and (b) LiDAR-fusion methods.
Table 5. Summary of 3D object tracking. Here ”I”, ”mvPC”, ”vPC”, ”pPC”, ”FrustumPC” stands for image, multiple view of point cloud,
voxelized point cloud, point cloud, Frustum point cloud respectively
Modality &
Category Model Architecture
Representation
DualBranch [72] mvPC Bbox growing method + multi-hypothesis extended Kalman filter
PV-RCNN [73] pPC & vPC Voxel-to-keypoint 3D scene encoding + keypoint-to-grid RoI feature abstraction
P2B [74] pPC Target-specific feature augmentation + 3D target proposal and verification
CenterPoint [76] pillar/vPC Map-view feature representation + center-based anchor-free head
LiDAR
SC-ST [80] pPC Siamese tracker(resemble the latent space of a shape completion network)
-Only
BEV-ST [81] mvPC Efficient RPN+Siamese tracker
PSN [82] pPC Siamese tracker(feature extraction + attention module + feature augumentation)
MLVSNet [83] pPC Multi-level voting+Target-Guided Attention+Vote-cluster Feature Enhancement
BAT [84] pPC Box-aware feature fusion + box-aware tracker
MSRT [85] I&pPC 2D object detector-Faster-RCNN+3D detector-Point RCNN
LiDAR MS3DT [86] I&mvPC Detection proposals+proposals matching&scoring+linear optimization
-Fusion Complexer-YOLO [87] I&vPC Frame-wise 3D object detetcion+novel Scale-Rotation-Transalation score
F-Siamese Tracker [88] I&FrustumPC Double Siamese network
6. 3D SEGMENTATION
3D Segmentation methods can be classified into semantic segmentation and instance segmentation, which are
both crucial for scene understanding of autonomous driving. 3D Semantic segmentation focuses on per-point
semantic label prediction so as to partition a scene into several parts with certain meanings (i.e., per-point
class labels), while 3D instance segmentation aims at finding the edge of instances of interest (i.e., per-object
masks and class labels). Since Kirillov et al. [89] first came up with the concept “panoptic segmentation” that
combines semantic segmentation and instance segmentation, several works [90,91] inspired by this concept have
been published recently, which build architectures for panoptic segmentation of point cloud. This section
specifically focuses on research concerning both 3D semantic segmentation and 3D instance segmentation