Recognizing Scenes from Novel Viewpoints

Code release for our paper

Recognizing Scenes from Novel Viewpoints

Shengyi Qian, Alexander Kirillov, Nikhila Ravi, Devendra Singh Chaplot, Justin Johnson, David F. Fouhey, Georgia Gkioxari

[Project Page] [arXiv] [BibTeX]

Setup

The code is tested on pytorch 1.11.0, detectron2 0.6, and pytorch3d 0.6.2.

conda create -n viewseg python=3.8
conda activate viewseg

# install pytorch, pytorch3d and detectron2 
conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch
conda install -c conda-forge -c fvcore -c iopath fvcore iopath
conda install pytorch3d -c pytorch3d
pip install 'git+https://github.com/facebookresearch/detectron2.git'

# install other packages
pip install accelerate==0.4.0
pip install submitit
pip install hydra-core --upgrade --pre
pip install hydra-submitit-launcher --upgrade
pip install visdom
pip install trimesh
pip install packaging plotly imageio imageio-ffmpeg matplotlib h5py opencv-contrib-python

Our checkpoints can be downloaded at viewseg_checkpoints.tar.gz. Please put them under checkpoints. deeplab_v3_ade20k is the deeplabv3+ encoder pretrained on ade20k using detectron2.

Data Preprocessing

Our primary datasets are Hypersim and Replica. All datasets are put under DEFAULT_DATA_ROOT specfied here.

For Hypersim, please refer to the original Hypersim repo to download Hypersim dataset and put it under viewseg_data/hypersim_renderings. Our generated source/target pairs can be downloaded at hypersim.tar.gz. Optionally, we preprocess a detectron2-style hypersim_sem_seg dataset for semantic segmentation evaluation only at hypersim_sem_seg.tar.gz.

For Replica, pre-processed data can be downloaded at replica.tar.gz. Optionally, we preprocess a detectron2-style replica_sem_seg dataset for semantic segmentation evaluation only at replica_sem_seg.tar.gz.

The final structure of the directory should look like

- viewseg_data
  - hypersim
  - hypersim_renderings
  - hypersim_sem_seg
  - replica
  - replica_sem_seg

Test our pretrained model

# Render 2D rgb and semantic images
python test_viewseg.py --config-name=viewseg_replica_finetune test.mode='export_imgs' hydra/launcher=submitit_local data.render_size='[192, 256]'

# Generate 3D Visualizations
python test_viewseg.py --config-name=viewseg_replica_finetune test.mode='export_mesh' hydra/launcher=submitit_local data.render_size='[192, 256]'

# Generate video visualizations
python test_viewseg.py --config-name=viewseg_replica_finetune test.mode='export_video' hydra/launcher=submitit_local data.render_size='[192, 256]'

Training

We use launcher.py to launch training and evaluation on slurm-managed cluster. For example,

# train on hypersim
python launch.py --config viewseg_hypersim_source_4 --mode train --nodes 4 --gpus 8

# eval on hypersim
python launch.py --config viewseg_hypersim_source_4 --mode eval --gpus 8
python eval.py --config-name=viewseg_hypersim_source_4 test.mode='evaluation' test.split='val' hydra/launcher=submitit_local

# train on replica
python launch.py --config viewseg_replica_finetune --mode train --gpus 8

# eval on replica
python launch.py --config viewseg_replica_finetune --mode train --gpus 8
python eval.py --config-name=viewseg_replica_finetune test.mode='evaluation' test.split='val' hydra/launcher=submitit_local

However, if slurm is unavailable or you just want to train it on a single gpu, try

# training
python train_viewseg.py --config-name viewseg_replica_finetune hydra/launcher=submitit_local encoder.bn='FrozenBN'

# evaluation
python eval_viewseg.py --config-name viewseg_replica_finetune hydra/launcher=submitit_local encoder.bn='FrozenBN'
python eval.py --config-name=viewseg_replica_finetune test.mode='evaluation' test.split='val' hydra/launcher=submitit_local

encoder.bn='FrozenBN' changes the Deeplabv3+ encoder batch normalization layer from SyncBN to FrozenBN, since SyncBN has to be run with distributed training. It can also be set to BN.

License

The code is released under CC-BY-NC 4.0.

Citation

@article{qian2021recognizing,
  title={Recognizing Scenes from Novel Viewpoints},
  author={Qian, Shengyi and Kirillov, Alexander and Ravi, Nikhila and Chaplot, Devendra Singh and Johnson, Justin and Fouhey, David F and Gkioxari, Georgia},
  journal={arXiv preprint arXiv:2112.01520},
  year={2021}
}