search

facebookresearch / localrf

An algorithm for reconstructing the radiance field of a large-scale scene from a single casually captured video.
MIT License
956 stars 62 forks source link

Why use laplacian.var() as a weight in color loss? #50

Closed fzhiheng closed 4 months ago

fzhiheng commented 4 months ago

Thanks for your great work. I have a question about RBG loss. Why use laplacian.var() as the weight for the RBG loss and what is the benefit of doing so?

all_laplacian = [np.ones_like(img[..., 0]) cv2.Laplacian(cv2.cvtColor((img255).astype(np.uint8), cv2.COLOR_RGB2GRAY), cv2.CV_32F).var() for img in all_rgbs]

loss = 0.25 ((torch.abs(rgb_map - rgb_train)) loss_weights) / loss_weights.mean()

Looking forward to your reply.

ameuleman commented 4 months ago

Hi, We use this scaling to reduce the supervision of blurier images. We did not formally evaluate its impact.