InfiniteNature-PyTorch

Unofficial PyTorch implementation of this ICCV 2021 Oral paper: Infinite Nature: Perpetual View Generation of Natural Scenes from a Single Image, developed by Yuan Shen. Check their project page for details: https://infinite-nature.github.io/.

Check our our NeurIPS 2022 SGAM paper to see how we generate consistent, realistic and large-scale 3D scene with more complicated scene objects, like building, than natural scene objects.

Colab Demo powered by our PyTorch model

To showcase the performance of this PyTorch codebase, I provide a similar colab demo presented in the original InfiniteNature in this Colab link. Enjoy!

Pre-trained Checkpoints

Download the converted Pytorch weight from their tensorflow checkpoint from this link and save in project root directory.

Environment

This codebase runs successfully in Python 3.7.13 and Ubuntu 22.03. First create your conda environment and then run

pip install -r requirement.txt
pip install torch==1.12.1+cu102 torchvision==0.13.1+cu102 -f https://download.pytorch.org/whl/torch_stable.html
pip install pytorch_lightning==1.5.10

BUILDING TF Mesh Renderer (Copy from their original codebase Readme.md)

We use the differentiable renderer from [here] (https://github.com/google/tf_mesh_renderer). We use gcc to build the library instead of their Bazel instructions to make it compatible with Tensorflow 2.0. To download and build: source download_tf_mesh_renderer.sh

tf_mesh_renderer was originally built for Tensorflow < 2.0.0, however we have prepared a small patch which upgrades the functions we use to work in Tensorflow 2.2.0. This means that the other parts of tf_mesh_renderer are still version incompatible.

And make sure you add the following scripts at the top of your main script, e.g., modify line 4 of main.py to direct to tf_mesh_renderer

# Make sure dynamic linking can find tensorflow libraries.
# os.system('ldconfig ' + tf.sysconfig.get_lib()) # You can also do in the terminal, cuz it might require root access. e.g. sudo ldconfig xxx, xxx can be printed out in python console from tf.sysconfig.get_lib()

# Make sure python can find our libraries. It needs absolute path, relative path doesn't work for me
sys.path.append('/.../infinite_nature') 
sys.path.append('/.../infinite_nature/tf_mesh_renderer/mesh_renderer')

# Make sure the mesh renderer library knows where to load its .so file from.
os.environ['TEST_SRCDIR'] = '/.../infinite_nature'

Inference

To run their similar Colab unrolling but using Pytorch, you can run the following command with fixed pose and next pose.

    python main.py

Feel free to modify the next pose to have more interesting trajectory. We manually check our output image for one-step extrapolation is exactly the same as those in Tensorflow

Train on your own dataset

For each dataset, we need to first prepare a config file and then write a customized data loader. Make sure you follow our provided config and dataloader example for GoogleEarth-Infinite dataset. The training is wrapped in Pytorch-lightning framework

    python train.py

For details about GoogleEarth-Infinite dataset, please check our NeurIPS 2022 project SGAM: Building a Virtual 3D World through Simultaneous Generation and Mapping.