search

chenhsuanlin / spatial-transformer-GAN

ST-GAN: Spatial Transformer Generative Adversarial Networks for Image Compositing :eyeglasses: (CVPR 2018)
MIT License
334 stars 72 forks source link

npy files in indoor dataset #24

Closed MeNinaa closed 5 years ago

MeNinaa commented 5 years ago

Hi,

can you please explain to me why some of the npy files in the dataset have different lengths (this is test dataset): `>>> a=np.load('perturb.npy')

b=np.load('disocclude.npy') a.shape (8258, 120, 160, 3) b.shape (2826, 120, 160, 3) ` And what is the goal of using randIdx and randIdxGT during training

Thanks Best regards Maria

chenhsuanlin commented 5 years ago

Sorry for the late reply. The length of disocclude.npy corresponds to the number of objects/scenes, and the length of perturb.npy is the number of generated perturbations (in the 3D space) from these objects. Since there may be multiple generated perturbations per each objects, the length of perturb.npy (as well as perturb_mask.npy) is larger than disocclude.npy (as well as remove.npy, indiv_mask.npy, and idx_corresp.npy).

In the code, randIdx randomly samples from the pre-generated dataset of 3D-perturbed objects, and randIdxGT associates them to the indices of the corresponding rendered scene through idx_corresp.

MeNinaa commented 5 years ago

Thanks!