Questions about data preparation for testing and evaluation mode

[dh001111]

Thanks for the great work on this project!

Currently I'm preparing my own data set for testing, and I noticed that all the samples in the datasets that you provided, no matter the Breaking Bad or the Fractura dataset, are initialized to a assembled state. But what if I have some fractured pieces that I only know that they belong to the same object but not sure what that object looks like? Does this kind of samples meets the model input requirements? I just generated a .hdf5 file using my own data and I found that "removal_masks" and "removal_order" are both zeros. I guess it is because all the pieces in my samples are just initialized in random places. Would this affect the model performance?

What's more, I also tried the eval mode of the model with the eval.sh file. It runed successfully but I want to know are there adjustable super parameters like steps and iterations so that I'm able to get similar results like when I use the demo?

Thanks in advance and looking forward to your reply!

[JDScript]

For the first part of questions, the answer is all these won't affact the performance.

While for the steps & iterations, you could pass:

++model.inference_config.max_iters=xxx
++model.inference_config.num_inference_steps=xxx

as indicated in our model config

[dh001111]

Thanks for your reply!

I have tried both the demo and the evaluation mode with my own data. I set all the super parameters that I can get to be identical, expect them to have the same, or similar output, but in the end I got totally different results.

This is the parameters I set in the demo:

I got a pretty nice result:

and I set the same parameters for eval.sh:

the result is different:

What could I miss here? I would really appreciate if you could have a look and provide a guidance. Thanks!

[JDScript]

If the demo is working, I believe that there should be an export button which you could export a json file from the website.

Try use the local visualization script to visualize that json file? You may need to change the name in the json dictionary to align with your HDF5 file.

[dh001111]

Actually I have tried your suggestion before and of course the output of the visualization script was the same as the demo result when using the .json file exported from the demo. But I have to rely on the local evaluation and visualization script if I want to fine tuning or use the framework in my own project. So I want to know what else differs the local script from the demo.