I found 2 problems with the visibility-based camera selection.
First, the method point_in_image in scene/vastgs/data_partition.py may be incorrect, camera.image_height and camera.image_width on line 403 & 404 and line 417 & 419 should be swapped.
Second, on dataset Mill-19, almost all of the points in the whole point cloud are included in one partition. The initial point cloud of dataset Mill-19 downloaded from https://vastgaussian.github.io/ is noisy and has many noise points under the ground, this makes the bounding box of points very large and may be the cause of this problem.
I found 2 problems with the visibility-based camera selection.
First, the method
point_in_image
inscene/vastgs/data_partition.py
may be incorrect,camera.image_height
andcamera.image_width
on line 403 & 404 and line 417 & 419 should be swapped.Second, on dataset
Mill-19
, almost all of the points in the whole point cloud are included in one partition. The initial point cloud of datasetMill-19
downloaded from https://vastgaussian.github.io/ is noisy and has many noise points under the ground, this makes the bounding box of points very large and may be the cause of this problem.