AbrahamYabo / Cascade-Zero123

61 stars 1 forks source link

How to train a custom dataset ? #3

Open Gaohhy opened 8 months ago

Gaohhy commented 8 months ago

Dear authors, Thank you for the great work. I can‘t fully download the dataset. Can you give me some advice on how to create my own dataset。what data structure should I process? Thanks

AbrahamYabo commented 8 months ago

Thanks for your interest. Firstly, the dataset of Zero123's Objaverse Renderings data from “https://tri-ml-public.s3.amazonaws.com/datasets/views_release.tar.gz” can be unzipped at any time, even if it is not completely downloaded. Just check and delete the sub-folders that images or npy have not been fully downloaded. Secondly, if you want to create own dataset, you can follow the setting of Zero 1-to-3. What's more, you can also rebuild the code of dataset.py to import your own dataset.

Gaohhy commented 8 months ago

Thanks very much for your reply!

xdobetter commented 5 months ago

Thank you very much for your work. I made my data set, but I found that my pose (.npy) was wrong. My npy obtained was set to the c2w matrix under NeRF, and then numerically, I found that my result is not correct compared with the npy of the original zero123 dataset, so I want to know if my pose setting method is correct. Could you give me some suggestions? Thank you very much.

Some intermediate results of training are as follows

my dataset my_dataset my npy file image

zero123 dataset origin_dataset

the npy file image

AbrahamYabo commented 5 months ago

Thank you very much for your work. I made my data set, but I found that my pose (.npy) was wrong. My npy obtained was set to the c2w matrix under NeRF, and then numerically, I found that my result is not correct compared with the npy of the original zero123 dataset, so I want to know if my pose setting method is correct. Could you give me some suggestions? Thank you very much.

Some intermediate results of training are as follows

my dataset my_dataset my npy file image

zero123 dataset origin_dataset

the npy file image

https://github.com/MattWallingford/ObjaVerse_PixelNerf/blob/master/src/data/objaverse.py This repo provide the pose transition from Objaverse(zero 1-to-3 pose coordinate) to Nerf(Blender pose coordinate). You can tried to reverse above operations.