lijiaman / omomo_release Goto Github PK

Hello,
I'm curious about the representation of the BPS (before being introduced to the diffusion model) in the format [x, y, z, x, y, z, ...] of size 1024*3. Specifically, I'd like to know if these values are normalized using z-score, min-max, or another method. Additionally, I'm wondering how missing values, such as 'nan' or 'none', from the BPS deltas are handled due to potential mismatches in some BPS points.

thanks

Hey,
Thank you for the amazing work. I would like to know if you are intending to release the raw data and the preprocessing code (getting bps representation)?

Thank you!

Hi jiaman,

Thank you for your excellent work.

I have a question regarding the dataset preprocessing code, about the variable 'obj_com_pos' mentioned in hand_foot_dataset.py L423. Although it seems that 'obj_com_pos' isn't used in the final training process, where com_pos, encoding the object's center using the mean vertices of object geometry, I wonder what is ‘obj_com_pos’ in L423 refer to?
Besides, i'm curious about the impact of various preprocessing steps, such as 'face front', start at xy=0. Do these preprocessing influence the performance?

Hello, thank you for opening up such excellent work. It seems that there is no text description in the dataset released by OMOMO. When will the FullBodyManipulation dataset with text and its corresponding CHOIS work be open source. Thank you for your reply

Hi, sorry to bother you. I want to know if there is any other way to generate visualization (if running code on the server)?

hi, thanks for the amazing work. I have trained as instructed for 400000 steps, however, when running test, the person mesh collapse, may I know to get the shared pretrained model checkpoint, how many training steps do we need?

Thank you for the fascinating work that you did!
I just wanted to know if there is the possibility of using the pretrained models on some of my videos!

Hello, thanks for your excellent work. I run the testing stage using the pretrained models you offered (It has been running for about three hours now), I wonder how much time does the testing stage take? The feedback I received is that Contact precision, TP + FP == 0!!, which means prediction no contact.

Hi,

Thanks for making this great work public. I'm checking the BPS representation of the objects and have a few questions.

It seems you use BPS to represent the whole trajectory of the object, encoding the absolute position of vertices. However, the BPS uses a fixed data grid for representation (a sphere with radius of 1). Is this done by using the custom_basis argument of bps_torch.encode? Does the argument offer the encoding in the object-centric coordinate space?

Thank you!