Indoor scene reconstruction with normal/depth priors

bennyguo / instant-nsr-pl

Neural Surface reconstruction based on Instant-NGP. Efficient and customizable boilerplate for your research projects. Train NeuS in 10min!

MIT License

856 stars 84 forks source link

Indoor scene reconstruction with normal/depth priors #18

Open bennyguo opened 2 years ago

bennyguo commented 2 years ago

[x] scannet dataset @pablovelaimmersed
[ ] nerf/neus training on scannet dataset
[ ] normal/depth priors

bennyguo commented 2 years ago

NeRF on scan2 after 100k iterations (left to right: gt, nerf render, depth, opacity):

https://user-images.githubusercontent.com/19284678/202697415-d3b6aefa-d3a1-40ee-9dfc-c658c94a2fa9.mov

pablovela5620 commented 2 years ago

Awesome, thanks for helping. I'll continue working on including a script to generate the depth/normal priors and integrate the relevant loss functions for both. I think the reason for bad quality results from Nerf is the lack of these geometric priors.