JacobiNeRF: NeRF Shaping with Mutual Information Gradient

Paper (CVPR 2023) | Project Page | Video

Xiaomeng Xu, Yanchao Yang, Kaichun Mo, Boxiao Pan, Li Yi, Leonidas Guibas

Dependencies

Main python dependencies are listed below:

• python>=3.8
• torch>=1.12.0
• cudatoolkit>=11.3

Requirements are listed in requirements.txt

Datasets

Download dataset

We mainly use Replica and ScanNet datasets for experiments, where we train a plain NeRF model and Jacobi-NeRF model on each 3D scene.

We also use pre-rendered Replica data provided by Semantic-NeRF.

Extract DINO feature

Extract per-pixel DINO feature from training images first. We use dino_vits8, layer 11 to create descriptors from. Please download the code of the DINO repo and run:

python extractor.py --model dino_vits8 --load_size 360 --image_path [image_path] --output_path [dino_path]

You can add a simple for loop in the DINO repo to extract DINO features for all images at once.

Dataset directory

Set dataset_dir in config_file./configs/replica_room0_config.yaml), structured as follows.

  dataset_dir
  ├── rgb 	# RGB images
  ├── semantic_class    # ground truth semantic segmentation labels
  ├── dino  # DINO features
  ├── traj_w_c.txt  # camera parameters
  └── ...

Running code

plain NeRF training

• Set N_iters in config_file./configs/replica_room0_config.yaml as 200000.

python main.py --config_file ./configs/replica_room0_config.yaml --save_dir [save_dir] --training_mode

Jacobi-NeRF training

• Set N_iters in config_file./configs/replica_room0_config.yaml as 10000.

python main.py --config_file ./configs/replica_room0_config.yaml --load_dino --load_on_cpu --feature_dim 384 --save_dir [save dir] --ckpt_path [plain nerf ckpt path] --contrastive_2d --contrastive_starting_step 0 --contrastive_step 1 --wgt_img 1 --wgt_contrastive 1e-2 --wgt_gradient 1e-2 --rgb_layer --pos_threshold 0.8 --neg_threshold 0.8 --adaptive_threshold --gt_feature_sim --gradient_norm_loss --N_rays 64 --contrastive_abs --training_mode

3D label propagation

• sparse setting

python main.py --config_file ./configs/replica_room0_config.yaml --sparse_views --sparse_ratio 0.995 --label_propagation --partial_perc 0 --load_saved --propagate_3d --ckpt_path [J-NeRF ckpt path] --save_dir [save_dir] --rgb_layer --mean_gradient

• dense setting

python main.py --config_file ./configs/replica_room0_config.yaml --sparse_views --sparse_ratio 0.995 --propagate_3d --ckpt_path [J-NeRF ckpt path] --save_dir [save_dir] --rgb_layer --adaptive_selection --num_comb 20 --num_iters 5 --train_agg

2D label propagation

• sparse setting

python main.py --config_file ./configs/replica_room0_config.yaml --sparse_views --sparse_ratio 0.995 --label_propagation --partial_perc 0 --load_saved --propagate_2d --ckpt_path [J-NeRF ckpt path] --save_dir [save_dir] --rgb_layer --mean_gradient

• dense setting

python main.py --config_file ./configs/replica_room0_config.yaml --sparse_views --sparse_ratio 0.995 --propagate_2d --ckpt_path [J-NeRF ckpt path] --save_dir [save_dir] --rgb_layer --adaptive_selection --num_comb 20 --num_iters 5 --train_agg

Acknowledgements

Thanks nerf, nerf-pytorch and Semantic-NeRF for nice implementation of NeRF and scripts in processing datasets.

Citation

If you find JacobiNeRF useful for your work, please cite:

@inproceedings{xu2023jacobinerf,
  title={JacobiNeRF: NeRF Shaping with Mutual Information Gradients},
  author={Xu, Xiaomeng and Yang, Yanchao and Mo, Kaichun and Pan, Boxiao and Yi, Li and Guibas, Leonidas},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={16498--16507},
  year={2023}
}

Contact

If you have any questions, please contact [email protected].