vye16 / shape-of-motion

MIT License
801 stars 57 forks source link

Rendering Quality is not good #8

Closed DevikalyanDas closed 3 months ago

DevikalyanDas commented 3 months ago

Hello @vye16 , thanks for your amazing work. I tried to run code following the steps (both for preprocessing and training) on a cat video from Banmo. I tried to obtain the rendering video using ''lemniscate'' trajectory but the quality is not good. Here is the video. I am not sure what is going wrong?

https://github.com/user-attachments/assets/06d852a6-2efe-4666-b248-d4f2b0bac6ed

vye16 commented 3 months ago

Hi, this can happen if the rendering trajectory is too far away from the training views, and depending on how the training cameras and depth maps look. A better way to debug is to run the rendering script python run_rendering.py --work-dir [WORK_DIR] to interactively view the trained model.

CTouch commented 2 months ago

@DevikalyanDas Did you get the video using run_video.py? I am trying som on nvidia-long dataset. I have already done the preprocess and train, got the custom dataset and ckpt. I am trying to get the rendering videos. Could you please tell me how did you made it? Thank you!

DevikalyanDas commented 2 months ago

@CTouch Yes I got the video using run_video.py. As far as I remember, the script was already working except for some changes in the dataloader for running the custom videos and you also have to provide the type of trajectory that you want to perform during rendering as commandline arguments. I think if the script is working for training the custom videos then it will work for rendering also.

CTouch commented 2 months ago

@CTouch Yes I got the video using run_video.py. As far as I remember, the script was already working except for some changes in the dataloader for running the custom videos and you also have to provide the type of trajectory that you want to perform during rendering as commandline arguments. I think if the script is working for training the custom videos then it will work for rendering also.

Got it! Thank you!