Trying 4DMOS on already aligned chunks

Hi Benemer,

I have several LAS files which are already registered in a global UTM-zone35N coordinate system. To use 4DMOS, I create chunks of my data with 250k points in each chunk with no overlap and convert it to bin format. Since the data is aligned, there is no pose information, so I used pretrained model D=5_scans_no_poses.ckpt to run the inference. Unfortunately, the results are not good at all. See attached images. I even offset the las files before chunking to the origin=[0, 0, 0] to make sure coordinates are not very large but it didn’t help. Do you have any suggestions? How can I create poses for aligned data and use other pretrained models? Or that’s not a problem at all?

Image of three consecutive chunks.

Results of the predictions with no poses. Red points are predicted MOS.

Expected results

[benemer]

You can try using our receding horizon strategy and fuse multiple predictions to get a better result.

[shayannikoohemat]

Many thanks for your feedback.