Misalignment between 3d mesh and 2d image

[tansor]

Hi, @Wei-Chen-hub @wqyin, I've visualized the MPII dataset to check the 3d mesh alignment with 2d image. However, the result shows that there is some misalignment even on ground-truth. Can you help me to check if I'm wrong?

MPII/data/images/071238764.jpg

MPII/data/images/049242269.jpg

from config import cfg
from utils.human_models import smpl_x
from utils.vis import render_mesh
import cv2

dataset = MPII(transform, 'train')

idx = 3
item = dataset[idx]  # return inputs, targets, meta_info, smplx_mesh_cam_orig
cam_param = dataset.datalist[idx]['smplx_param']['cam_param']
mesh = item[3]  # smplx_mesh_cam_orig
img = cv2.imread(dataset.datalist[idx]['img_path'])
vis_img = render_mesh(img, mesh, smplx.face, cam_param, mesh_as_vertices=True)
cv2.imwrite('render_mesh.png', vis_img)

And also, I wanna know if there are some datasets, that are well fit between 3d mesh and 2d image?
Thanks a lot !!

similar issue: #45 #54

[gaogehan]

I also found alignment issues when testing the COCO whole-body training set using the corrected model. Is there a solution to this problem?