3D reconstruction of custom dataset

[njfugit]

Hello, I am reconstructing a scene using an existing pretrained model. It is found that point cloud stacking occurs during the point cloud fusion process. In addition, I output the predicted depth of each frame of image to generate a single frame point cloud. Indeed, the point cloud of each frame is also stacked. I want to know how to solve this situation? Which processing link has a problem, my pose comes from a VIO SLAM system.
The point cloud of a single frame will overlap. This is the result of displaying three frames of point clouds together:

The fused point cloud stacking is more serious: