search

skhu101 / GauHuman

Code for our CVPR'2024 paper "GauHuman: Articulated Gaussian Splatting from Monocular Human Videos"
Other
339 stars 29 forks source link

Initialization with a picture instead of a vedio #51

Open ccxi1008 opened 3 weeks ago

ccxi1008 commented 3 weeks ago

Thanks for your work! I'm trying to use one picture instead of a vedio to get the initialized 3D Gaussians in pose space with respect to SMPL and LBS model ，and still optimize the Guassians using the follow-up control process similar to the article. I'm wondering that:

  1. if the 2 MLPs mentioned in the article are dedicated for each input data ? (Am I need to train dedicated MLPs for every input each time？）
  2. What part of the code can I modify to change the input video to a single image？(Including what should I have for every item in my picture dataset)
  3. Should my data need ground truth SMPL for every picture for initialization when training? Or I need SMPL for every item just for evaluation? Thanks！
skhu101 commented 2 weeks ago

Hi, thanks for your interest in our work. Do you hope to train a 3D Gaussian model with only one image? If this is the case, 1). you need to train 2 MLPs mentioned in the article. 2). You can refer to scene/dataset_readers.py to change the input from a video to a single image. 3). You need ground truth SMPL for training and testing data.