search

ActiveVisionLab / nope-nerf

(CVPR 2023) NoPe-NeRF: Optimising Neural Radiance Field with No Pose Prior
https://nope-nerf.active.vision/
MIT License
372 stars 30 forks source link

How to reproduce two-stage training？ #25

Open KingofLondon opened 10 months ago

KingofLondon commented 10 months ago

Great work. As you mention in your paper sec 4.3, you use both one-stage training and two stage training，and two stage training outperforms one-stage training. But I dont see two stage training in the released code. Can you tell me how to run two stage training with this code?

bianwenjing commented 8 months ago

Hi, thanks for your interest in the project. For the two-stage training approach, you can train the first stage using the standard method. Subsequently, train the second stage by configuring cfg['training']['load_pose_dir'] to the checkpoint directory of the model trained in the first stage, set cfg['training']['load_ckpt_model_only'] to True and set other parameters according to NeRF training. Ensure that cfg['training']['load_dir'] is set to None so that the NeRF model in the initial stage is not loaded.