Closed sunwonlikeyou closed 1 year ago
You could revise the pose parameters here: https://github.com/zju3dv/animatable_nerf/blob/master/lib/datasets/tpose_pdf_pose_sequence_dataset.py#L99
Thank you for reply :) By the way, do i get very weired result when I use very different from training poses?
No. Our model generalizes to various human poses. Here are some results on difficult poses: https://zju3dv.github.io/animatable_sdf/
In your paper, and https://github.com/zju3dv/animatable_nerf/issues/46 Do I have to retrain novel poses?? and is it okay without RGB and masks? Because I asked this issue before reading about https://github.com/zju3dv/animatable_nerf/issues/46. And sdf_pdf model doesn't have 'novel_pose_bw' When I revised
# training the blend weight fields of unseen human poses
python train_net.py --cfg_file configs/aninerf_s9p.yaml exp_name aninerf_s9p_full resume False aninerf_animation True init_aninerf aninerf_s9p
this into my data,( actually that code for init iccv version) I got this messege
AttributeError: 'Network' object has no attribute 'novel_pose_bw'
The extended version does not require the re-training, while the ICCV version needs it.
Then, there is no problem just revising the config file
vertices: 'vertices' params: 'params'
these are something novel poses like amass or aist?? The first condition should be training pose and novel poses are on the same world.
I'm so sorry but Can I get a novel pose and pretrained model?? I mean some poses and pretrained avatars in https://zju3dv.github.io/animatable_sdf/
Hi, hello, have you achieved success?
First, Thank you for sharing codes!! I'd like to synthesize novel pose on extended model. Novel pose means like random smpl poses, e.g. aist++ datasets or AMASS. datasets.
There is a code to retrain blend weights of unseen poses. But There is no code to retrain the unseen poses for extended model and also that code seems like training for novel frames.
How can I synthesize totally unseen poses?