Open huster-wugj opened 4 weeks ago
Hi, I think it related to your initial point cloud. New points are added to be close to the initial points. You could try to use some pretrained depth model to predict a depth map for your input image, convert the depth to point cloud and then sample some points from it as for initialization.
Great job. But when i train on my own data, i found that the quality of pointclouds has not a hight quality, especially the road surface.![Screenshot from 2024-08-14 11-00-36](https://github.com/user-attachments/assets/32344237-1afa-4a27-ace7-fb3750b2bb11)