salingo / rpm-net Goto Github PK

Hi,
Thank you very much for releasing your repository. While reading your paper and the github, something caught my attention in the segmentation module. You used DBSCAN to group features into clusters (representing the motion parts). For that matter, you fixed epsilon hyper-parameter (The maximum distance between two samples for one to be considered as in the neighborhood of the other) to 10. If you don't mind I was wondering why such a choice? What is the tuning process you used to fix that hyper parameter?
Is there any link between the choice of this hyper-parameter and the densities of the features?

Thank you very much in advance.

I have trained the network for 120 epochs, and tried visualizing the frames output on the test data. Here are some example results (others that I've inspected look similar, with random bits getting moved):

Is this a sign that something has gone horribly wrong in the training, or is this something that is expected to happen?

The losses reported at the final epoch of training were as follows:

**** EPOCH 120 ****
2021-08-03 18:46:09.223722
EPOCH STAT:
mean mov loss: 1.437132
mean ref loss: 0.047632
mean disp loss: 0.118962
mean mov seg loss: 0.418661
mean part seg loss: 1.015106
mov seg acc: 0.984643
part seg err: 1.312428
Model saved in file: ../output/model_rpm_2021-08-02-15-04-37_all/ckpts/model.ckpt-120

Is a Titan Xp insufficient for training this model? How much VRAM does RPM-Net require? Are there recommended ways to reduce the memory usage with the least impact on quality?

Hi! Congratulations on the cool work with RPMNet. The paper also mentions the Mobility-Net for predicting joint parameters but I am unable to find it in the code base. Can you please point me to where I can get that?

Thanks!