Graham-Hall_Colmap Dataset (Splatfacto-mcmc)

169

1. Download the dataset from https://demuc.de/colmap/datasets/ 2. Train it with Nerfstudio 3. Settings were used: * splatfacto-mcmc * using bilateral-grid * 60k steps * 3 M splats * batch-size = 3 * ⁨⁨means_lr = 1.6e-05 and ⁨⁨opacity_reg = 0.001 * resize_factor = 4 (I resize it with ⁨⁨⁨cv2.resize(img, (width/4, height/4), interpolation=cv2.INTER_AREA))

Suggested

2 Comments

ichsan28958 months ago

@simonbethke Yeah, the quality of individual frames needs improvitation. It has many exposure changes. Bilateral grid helps a lot.

simonbethke8 months ago

DAMN! I thought this is a nice dataset. I downloaded it, fixed the exposure and posed it. However, there is almost NO image without motion-blur.... the quality of the individual frames is a pain

Loading comments...