Table 16.

Literature Analysis: Graph-based Representation for 3DOR.

	Point- GCNN [69]	RGNet [70]	HGNet [71]	S-AT GCN [72]
Detector Category	One-stage	Two-stage	Two-stage	Two-stage
Environment	Outdoor	Indoor	Indoor	Outdoor
Scenario	Object detection scenario from a LiDAR point cloud using Graph neural network	3D object proposal generation and relationship extraction scenario in point cloud using relation graph network	Raw point clouds processing scenario for direct 3D bounding box prediction.	Local geometrical feature extraction scenario
Advantage(s)	Detects multiple objects by predicting their category and shape in a single shot with auto registration mechanism	Extracts uniform appearance features by point attention pooling method Holds appearance and position relationship between 3D objects by building a relation graph	Learns semantics via hierarchical graph representation, Applies multi-level semantics by capturing the relationship of the points to detect 3D objects	FE layers boost the contrast ration of feature map and increase the 3D recognition (true positive) rate of the subsequent CNN for small and sparse objects
Limitation(s)	Does not maintain the accuracy with down sampled data for the hard and moderate levels	Gives poor performance for detecting thin objects	The ProRe module is not effective for object detection if object features had been adequately learned	Run-time speed drops with FE layers