NeRF++ TensorFlow Repository

This code repository aims to build NeRF++ with TensorFlow.

Original Repository: https://github.com/Kai-46/nerfplusplus

Requirements

• Install below via pip:
  • tensorflow-gpu
  • miniball
  • python-box
  • opencv-python

We used environment specified in requirements.txt. Use command below to install all the packages.

pip install -r requirments.txt

How to train

Training example (please refer to training_examples.sh):

python main.py ckpt/nerf_africa_nolratedecay --use_viewdirs --n_samples 128 --dataset_type nerfppdata --dataset_name africa --n_importance 256 --use_perturb --lrate_decay 50000 \
--rays_per_batch 1024 --gpus 0 --chunk_size 2048 --val_step 5000 --max_step 1000000 --no_batching \
--learning_rate 5e-4

How to evaluate

evaluate example (use the same network settings as training):

python evaluate.py ckpt/nerf_africa_nolratedecay/ckpt-250000 \
--model_type nerfpp --dataset_type nerfppdata --dataset_name africa \
--use_viewdirs --n_samples 64 --n_importance 128 \
--chunk_size 4096 \
--normalize_coordinates --gpus 0

Evaluate pretrained model example

Pretrained models available in nerf_checkpoints_export directory.

python evaluate.py nerf_checkpoints_export/lf_tnt/nerfpp_africa/ckpt-1 \
--model_type nerfpp --dataset_type nerfppdata --dataset_name africa \
--use_viewdirs --n_samples 64 --n_importance 128 \
--chunk_size 4096 \
--normalize_coordinates --gpus 0

Datasets

We provide links to download: NeRF Synthetic, DTU, LF, Tanks and Temples (T&T).

Dataset Directories

Please place each dataset as below:

• [WORKING DIRECTORY]
  • nerfdata
    • nerf_synthetic
      • lego
      • chair
      • ...
  • nerfppdata
    • africa
    • basket
    • scan65
    • scan106
    • scan118
    • ship
    • statue
    • tat_intermediate_M60
    • tat_intermediate_Playground
    • tat_intermediate_Train
    • tat_training_Truck
    • torch