Is it possible to push it to speed up 3D instance segmentation in this way?

[Neo-YH]

https://github.com/Neo-YH/YOLOXYZ
Can this be a method to accelerate 3D instance segmentation? Or does this method have little value compared with the existing neural network that processing point cloud? Dealing point clouds from different views might be out of time.

[Neo-YH]

I am trying to apply the deep learning method in industrial scenarios where both 2D and 3D cameras exist. YOLOX performs really well in object detection. As for point cloud processing, we are still using traditional methods such as matching shapes via PCL-supported libraries. So I am wondering if it is possible to combine 2D and 3D methods to do my project in a quick way. Recently, I am reading papers about Frustum PointNet and DGCNN, but not sure if they can handle the large point cloud in a quick way. So maybe it is possible to find the approximate box in a quick way, and then detailed segmentation will be implemented in a small area. Most of the time, the upper side won't be blocked in industrial areas.