search

pals-ttic / sjc

Score Jacobian Chaining: Lifting Pretrained 2D Diffusion Models for 3D Generation (CVPR 2023)
https://pals.ttic.edu/p/score-jacobian-chaining
Other
500 stars 15 forks source link

Confused about sampled poses #28

Open OrangeSodahub opened 10 months ago

OrangeSodahub commented 10 months ago

Hi, I fetched some generated poses by Poser() and some examples:

n_steps = 10
poser = Poser(H=128, W=256, fov=60., R=1.5)
Ks, poses, prompt_prefixes = poser.sample_train(n_steps)
print(Ks)
print(poses)

# Ks
[array([[ 287.60829222,    0.        , -127.5       ],
       [   0.        , -287.60829222,  -63.5       ],
       [   0.        ,    0.        ,   -1.        ]]), ...]
# poses
[[[-0.37924771 -0.74374956  0.55046142  0.82569212]
  [ 0.          0.59490358  0.80379707  1.20569561]
  [-0.92529518  0.3048382  -0.22561582 -0.33842373]
  [ 0.          0.          0.          1.        ]], ...]

So I was wondering why the intrinsic matrix has a negative value of fy, cx, cy, I think it is just because the speciality of the different coordinates systems of your definition? And if I replace the generation with some given poses (e.g. real Ks and poses captured in reality), it should also work.