the gvo res on SceneNN seems peculiar

[crazyMessi]

I hope this message finds you well. I wanted to express my gratitude for the excellent job you have done. Your work has been truly impressive.

I am particularly interested in understanding how your job works on the open model. In order to gain a better understanding, I took the liberty of running the following commands:

1. Train sceneNN's NGL:

   python 01_train_test_NGL.py --mode=train --gpu=0 --data_set=SceneNN

   Please note that I have already made the necessary modifications to the database root.

2. Train GVO using PCPNN: Unfortunately, I encountered some difficulties while attempting to train GVO using SceneNN. As a result, I decided to train GVO using PCPNN, following the default settings.

   python 02_run_GVO.py --mode=train --gpu=0

3. Test NGL model on SceneNN:

   python 01_train_test_NGL.py --mode=test --gpu=1 --data_set=SceneNN --ckpt_dir=231206_044902_SceneNN --ckpt_iter=40000

   The results of testing the NGL model on SceneNN were quite promising. The average angle difference was approximately 20-30, which aligns with the motion described in your paper.

4. Test GVO model (trained by PCPNN) on SceneNN:

   python 02_run_GVO.py --mode=test --gpu=3 --data_set=SceneNN --ckpt_dirs=231206_080034_GVO --ckpt_iters=850 --normal_init ./log/231206_044902_SceneNN/test_40000/pred_normal/

   However, when I tested the GVO model (trained by PCPNN) on SceneNN, the results were rather peculiar. I am uncertain whether I made any mistakes during the process.

I would greatly appreciate it if you could provide some guidance regarding the peculiar results obtained during the testing of the GVO model on SceneNN. Any insights or suggestions you may have would be invaluable.

Thank you in advance for your time and assistance.

[crazyMessi]

I tried again and the results seem more reasonable. Although there are still some points with incorrect orientation, the GVO worked.