Open shangguan9191 opened 4 years ago
Hi @shangguan91, Thanks for your interest. That's an excellent question, we are thinking on that direction as well. Point-GNN's classification header already categorizes each point as the type of the object it belongs to. From this perspective, a segmentation task only needs the classification header to predict point label and we can remove the localization header. Now, the tricky part is the computational cost. A graph with raw point cloud and edges connecting them would be large. The network would be very slow. Currently, we downsample the point cloud to reduce the size of the graph. This is fine with object detection because we can still get a bounding box as long as there are some points on the object. But for segmentation, we want the label for all points. To make it efficient, we probably need to add a "pyramid" structure.
Dear Shi,
i am researching methods applying on 3d data, Is Point-GNN suitable for classification and segmentation Task?