My own dataset visualization

[MayuOshima]

Hello, thanks for your great work!

Actually, I have used your code to train my own dataset and saved the dict.
Now I want to render the predicted meshes.
How can I get the renderings using the saved 'verts', 'pose', 'betas', 'cam', 'trans'?
I've got the SMPL model and read your code visualization.py. However, it seems you don't use the 'cam' as cam_param.
If I do what visualization.py does, I can't get a correct rendered image.

So could you please tell me how to get the correct rendered images using saved dict information?

[Arthur151]

To visualize the mesh on image, you just need to run visualize_renderer_verts_list with right inputs.

The vertex are just verts in npz.

The estimated translation of meshes are cam_trans.
I just submit a commit to master to save cam_trans into npz via this line and this line.

You just need to collect the npz results into tow lists and put them into visualize_renderer_verts_list. For instance, the code might like

import numpy as np
import torch
import cv2

results = np.load('path/to/npz',allow_pickle=True)['results'][()]
for img_path in results:
    image_results = results[img_path]
    org_imgs = [cv2.imread(img_path)]
    per_img_verts_list = [torch.from_numpy(image_results[person_id]['verts']).float() for person_id in image_results]
    mesh_trans = [torch.from_numpy(image_results[person_id]['cam_trans']).float() for person_id in image_results]
    rendered_imgs = self.visualize_renderer_verts_list(per_img_verts_list, images=org_imgs, trans=mesh_trans)

Just like that. Enjoy your day with ROMP~
Best,
Yu

[MayuOshima]

Thanks for your quick reply!
I have tried your code and it works!