sparf

SPARF: Neural Radiance Fields from Sparse and Noisy Poses

This is not an officially supported Google product.

This repository contains the code for the paper: SPARF: Neural Radiance Fields from Sparse and Noisy Poses. In CVPR, 2023 (Highlight).

Authors: Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari

[arXiv preprint] [Website] [Youtube teaser]

Our approach SPARF produces realistic novel-view rendering given as few as 2 or 3 input images, with noisy camera poses. We add two novel constraints into the pose-NeRF optimization: the multi-view correspondence loss and the depth-consistency loss.

Please contact Prune Truong ([email protected]) if you have any questions!

We provide PyTorch code for all experiments: BARF/SPARF for joint pose-NeRF training, NeRF/SPARF when considering fixed ground-truth poses as input.

---

Citation

If you find our code useful for your research, please cite

@inproceedings{sparf2023,
  title={SPARF: Neural Radiance Fields from Sparse and Noisy Poses},
  author = {Truong, Prune and Rakotosaona, Marie-Julie and Manhardt, Fabian and Tombari, Federico},
  publisher = {{IEEE/CVF} Conference on Computer Vision and Pattern Recognition, {CVPR}},
  year = {2023}
}

---

Installation

This code is developed with Python3 (python3) with Cuda 11.3. All models were trained on single A100 (or V100) GPU with 20GB of memory. This code does NOT support multi-GPU training. Please use the following command for installation.

• It is recommended to use Anaconda to set up the environment. Install the dependencies and activate the environment sparf-env with

conda create -n sparf-env python=3
conda activate sparf-env
pip install -r requirements.txt

For the specific versions of the packages used to develop this model, run the following instead.

pip install -r requirements_w_version.txt

• Make sure to the dependencies of the submodules are downloaded. Everything should be included in the provided requirement file, except for cupy: We use PDC-Net to extract correspondences between the views, for which CuPy is a required dependency. It can be installed using pip install cupy or alternatively using one of the provided binary packages as outlined in the CuPy repository. For another CUDA version, change accordingly.

pip install cupy-cuda113 --no-cache-dir 

• Initialize the external submodule dependencies with

git submodule update --init --recursive
git submodule update --recursive --remote

• Optional: The submodule contains lots of unnecessary files. You can suppress them by running

bash third_party/remove_unused_files.sh

• Create source/admin/local.py by running the following command and update the paths to the datasets and workdirs. We provide an example admin/local_example.py where all datasets are stored in data/.

python -c "from source.admin.environment import create_default_local_file; create_default_local_file()"

• Download the pre-trained model of PDC-Net here. You will need to change the path to the pre-trained model in train_settings/default_config.py cfg.flow_ckpt_path (L.190).

Run the following command

python third_party/test_pdcnet_installation.py

You should obtain the following image as third_party/test_pdcnet.png

---

Datasets

• Images: We use the DTU dataset, processed by PixelNeRF, where the images are processed and resized to 300 x 400. Download the data here.

• Mask Data: For evaluation, we report also masked metrics. For this, we use the object masks provided by DVR, IDR and RegNeRF. RegNeRF provides the full mask data (for the test splits), which you can download here.

• Ground-truth depth maps: For evaluation, we report the depth error. For this, we download the depth maps. They are extracted from MVSNeRF.

The LLFF real-world data can be found in the NeRF Google Drive. For convenience, you can download them with the following script: (under this repo)

gdown --id 16VnMcF1KJYxN9QId6TClMsZRahHNMW5g # download nerf_llff_data.zip
unzip nerf_llff_data.zip
rm -f nerf_llff_data.zip
mv nerf_llff_data data/llff

You can download the replica dataset with the following script:

# you can also download the Replica.zip manually through
# link: https://caiyun.139.com/m/i?1A5Ch5C3abNiL password: v3fY (the zip is split into smaller zips because of the size limitation of caiyun)
wget https://cvg-data.inf.ethz.ch/nice-slam/data/Replica.zip
unzip Replica.zip

---

I. Running the code

Training: Quick start

The installation should have generated a local configuration file "admin/local.py". In case the file was not generated, run python -c "from source.admin.environment import create_default_local_file; create_default_local_file()"to generate it.

• Next, set the paths to the training workspace self.workspace_dir, i.e. the directory where the model weights and checkpoints will be saved.

• Set the path to self.tensorboard_dir, where the tensorboard files will be stored

• Set the path to self.log_dir, where the logs will be stored

• Set the path to self.eval_dir, where the test results will be stored.

• Also set the paths to the datasets you want to use (and which should be downloaded beforehand, see above).

• Set the path to the PDC-Net pre-trained model (previously downloaded) in train_settings/default_config.py cfg.flow_ckpt_path.

If all the dependencies have been correctly installed, you can train/evaluate a network using the run_trainval.py script in the correct conda environment. All checkpointing, logging and saving metrics should be done automatically.

conda activate sparf-env
## Selecting <train_module> <train_name> <nbr_input_views> <scene_name>
python run_trainval.py <train_module> <train_name> --train_sub <nbr_input_views> --scene <scene_name>

Here, train_module is the sub-module inside train_settings and train_name is the name of the train setting file to be used.

The snapshots along with a video of the RGB and depth renderings will be saved in the directory workspace_dir/<train_module>/subset_<nbr_input_views>/<scene_name>/<train_name>.

The corresponding tensorboard file will be stored in tensorboard_dir/<train_module>/subset_<nbr_input_views>/<scene_name>/<train_name>.

where workspace_dir and tensorboard_dir were set in the source.admin.local.py file.

Running this command directly computes results on the test set at the end of training, which will be stored in eval_dir/dataset_name/<train_module>/subset_<nbr_input_views>/<scene_name>/<train_name>. If you wish to recompute them, or re-generate the video rendering, run:

## Selecting <train_module> <train_name> <nbr_input_views> <scene_name>
## to regenerate the test metrics 
python run_trainval.py <train_module> <train_name> --train_sub <nbr_input_views> --scene <scene_name> --test_metrics_only True 

## to regenerate the video of the renderings
python run_trainval.py <train_module> <train_name> --train_sub <nbr_input_views> --scene <scene_name> --render_video_only True 

Training: Example

The configs are found in train_settings

For example, you can train using the included default SPARF settings for joint pose-nerf training on DTU, starting from initial noisy poses (3 views) by running:

python run_trainval.py joint_pose_nerf_training/dtu sparf --train_sub 3 --scene scan82

All the snapshots will be stored in the directory workspace_dir/joint_pose_nerf_training/dtu/subset_3/scan82/sparf. All the tensorboard will be stored in tensorboard_dir/joint_pose_nerf_training/dtu/subset_3/scan82/sparf.

Running this command directly computes results on the test set at the end of training, which will be stored in eval_dir/dtu/joint_pose_nerf_training/dtu/subset_3/scan82/sparf.

Visualizing the results

We have included code to visualize the training over TensorBoard. The TensorBoard events include the following:

• SCALARS: the rendering losses and PSNR over the course of optimization. For joint pose-NeRF optimization, the rotational/translational errors with respect to the given poses are also computed. The validation metrics are also logged.
• IMAGES: visualization of the RGB images and the RGB/depth rendering. For joint pose-NeRF optimization, we also plot the optimized and ground-truth poses.

At the end of the training, a video of the RGB and depth renderings is created, and saved with the snapshots. The poses used for the rendering are jitters of the optimized (or ground-truth) input poses.

Evaluation of pre-trained models

The computation of the test metrics is included at the end of the training. But it must correspond to an existing train_settings file. To evaluate any given pre-trained models (associated with its options.yaml file), run:

## Selecting <ckpt_dir> <out_dir> <expname>
python eval.py --ckpt_dir <ckpt_dir> --out_dir <out_dir> --expname <expname> --plot True

This will save the metrics file at <out_dir>/<expname>.json. if --plot is True, figures of the renderings are saved in <out_dir>/<expname>/. <ckpt_dir> is the path to a directory containing a checkpoint (the latest will automatically be loaded) and an option file named options.yaml.

We provide pre-trained models for some of the experiments and some of the scenes here.

To make sure the code is running fine, for joint_pose_nerf_training/dtu/scan82, you should get similar metrics (with test-time optimization):

	Rot. error	Trans. error	PSNR (masked)	SSIM (masked)	LPIPS (masked)	Depth err.
SPARF (Ours)	0.70	0.0097	17.58 (17.36)	0.82 (0.91)	0.19 (0.07)	0.15

Reproducing results from the publication

In Tab. 4 of the main paper, we present results of the joint pose-NeRF optimization on DTU, starting from 3 images with noisy poses. To reproduce those results, run these commands for each of the 15 scenes. The results presented are the average over all scenes.

## <SCENE> is specific to datasets
## DTU (<SCENE>={'scan8', 'scan21', 'scan30', 'scan31','scan34', 'scan38','scan40','scan41','scan45','scan55','scan63','scan82','scan103','scan110','scan114'})

## SPARF
python run_trainval.py joint_pose_nerf_training/dtu sparf --scene=<SCENE> --train_sub 3

## SPARF without depth-consistency loss
## Note that the depth-consistency loss adds training time and only leads to 
## a minor improvement. We therefore also include a version without the depth
## consistency loss 
python run_trainval.py joint_pose_nerf_training/dtu sparf_wo_depth_cons_loss --scene=<SCENE> --train_sub 3

## Baseline BARF
python run_trainval.py joint_pose_nerf_training/dtu barf --scene=<SCENE> --train_sub 3

To make sure the code is running fine, for scene='scan82', you should get similar metrics (with test-time optimization):

	Rot. error	Trans. error	PSNR (masked)	SSIM (masked)	LPIPS (masked)	Depth err.
BARF	13.28	0.41	7.79 (6.11)	0.45 (0.77)	0.67 (0.28)	1.53
SPARF (Ours)	0.70	0.0097	17.58 (17.36)	0.82 (0.91)	0.19 (0.07)	0.15

Scene 'scan30' is an example of failure cases, because no correspondences can be reliably extracted.

We provide pre-trained models of our SPARF with 3 input views for all 15 scenes here. To evaluate them, run

## Selecting <ckpt_dir> <out_dir> <expname>
python eval.py --ckpt_dir <ckpt_dir> --out_dir <out_dir> --expname <expname> --plot True

In Tab. 5 of the main paper, we present results of the joint pose-NeRF optimization on LLFF, given 3 images with identity initial poses. To reproduce those results, run these commands for each of the 8 scenes. The results presented are the average over all scenes.

## <SCENE> is specific to datasets
## LLFF (<SCENE>={'orchids', 'horns', 'trex', 'fern', 'flower', 'leaves', 'room', 'fortress'})

## SPARF
python run_trainval.py joint_pose_nerf_training/llff sparf --scene=<SCENE> --train_sub 3

## SPARF without depth-consistency loss
python run_trainval.py joint_pose_nerf_training/llff sparf_wo_depth_cons_loss --scene=<SCENE> --train_sub 3

## Baseline BARF
python run_trainval.py joint_pose_nerf_training/llff barf --scene=<SCENE> --train_sub 3

To make sure the code is running fine, for scene='horns', you should get similar metrics (with test-time optimization):

	Rot. error	Trans. error	PSNR	SSIM	LPIPS
BARF	5.53	0.326	14.34	0.34	0.54
SPARF (Ours)	0.027	0.002	18.94	0.61	0.33

In Tab. 6 of the main paper, we present results of the joint pose-NeRF optimization on Replica, starting from 3 images with initial poses obtained by COLMAP with PDC-Net matches. To reproduce those results, run these commands for each of the 7 scenes. The results presented are the average over all scenes.

## <SCENE> is specific to datasets
## Replica (<SCENE>={'room0', 'room1', 'room2',  'office0', 'office1', 'office2', 'office3'})

## SPARF
python run_trainval.py joint_pose_nerf_training/replica sparf --scene=<SCENE> --train_sub 3

## SPARF without depth-consistency loss
python run_trainval.py joint_pose_nerf_training/replica sparf_wo_depth_cons --scene=<SCENE> --train_sub 3

## Baseline BARF
python run_trainval.py joint_pose_nerf_training/replica barf --scene=<SCENE> --train_sub 3

To make sure the code is running fine, for scene='office0', you should get similar metrics (with test-time optimization):

	Rot. error	Trans. error	PSNR	SSIM	LPIPS	Depth err.
BARF	5.37	0.28	22.33	0.72	0.30	0.59
SPARF (Ours)	0.58	0.010	28.38	0.90	0.13	0.36

In Tab. 7 of the main paper, we present results of NeRF-based approaches, trained with fixed ground-truth camera poses. To reproduce those results, run these commands for each of the 15 scenes. The results presented are the average over all scenes.

## <SCENE> is specific to datasets
## DTU (<SCENE>={'scan8', 'scan21', 'scan30', 'scan31','scan34', 'scan38','scan40','scan41','scan45','scan55','scan63','scan82','scan103','scan110','scan114'})

## SPARF
python run_trainval.py nerf_training_w_gt_poses/dtu sparf --scene=<SCENE> --train_sub 3


## Baseline NeRF
python run_trainval.py nerf_training_w_gt_poses/dtu nerf --scene=<SCENE> --train_sub 3

To make sure the code is running fine, for scene='scan82', you should get similar metrics:

	PSNR (masked)	SSIM (masked)	LPIPS (masked)	Depth err.
NeRF	4.57 (5.36)	0.28 (0.77)	0.28 (0.32)	1.35
SPARF (Ours)	18.42 (21.71)	0.87 (0.95)	0.16 (0.04)	0.24

In Tab. 7 of the main paper, we present results of NeRF-based approaches, trained with fixed ground-truth camera poses. To reproduce those results, run these commands for each of the 8 scenes. The results presented are the average over all scenes. Almost the same results are obtained using hierarchical sampling, or with only a coarse sampling.

## <SCENE> is specific to datasets
## LLFF (<SCENE>={'orchids', 'horns', 'trex', 'fern', 'flower', 'leaves', 'room', 'fortress'})

## SPARF
python run_trainval.py nerf_training_w_gt_poses/llff sparf --scene=<SCENE> --train_sub 3


## Baseline NeRF
python run_trainval.py nerf_training_w_gt_poses/llff nerf --scene=<SCENE> --train_sub 3

To make sure the code is running fine, for scene='horns', you should get similar metrics:

	PSNR (masked)	SSIM (masked)	LPIPS (masked)
NeRF	13.21	0.26	0.61
SPARF (Ours)	19.39	0.64	0.29

---

II. Using a pre-trained model

You can find our pre-trained models (from running this repo), following the same structures as our train_settings files, here. After downloading the checkpoints, you can evaluate the model and render test images as indicated above.

---

III. Downloading model predictions

If you don't want to run the code, you can also directly download the renderings of our models for the test set poses using the following link: https://drive.google.com/drive/folders/1lHryExsutZsbcKJlzO7QKM34YlUwKSSx?usp=share_link. The structure follows that of the train_settings files. They correspond to the renderings used in the paper, so they will be a bit different compared to the ones obtained when running the provided pre-trained models (for which the training was done with this codebase).

---

IV. Codebase structure

For details on the dataloader and how to use your own, refer to this doc.

The framework in source/ consists of the following sub-modules.

• training:
  • core: the losses, sampling strategy and metric files
  • engine: Classes for main training routine
  • base.py: Base class for the NeRF/pose trainer. All the others inherit from it.
  • nerf_trainer.py: The main engine for training a NeRF model, considering fixed ground-truth poses. Only the NeRF model is optimized. It inherits base.py.
  • joint_pose_nerf_trainer.py: The main engine for training of the NeRF model jointly with refining initial noisy camera poses. This corresponds to SPARF's main use case. Optimized are the NeRF model and the camera pose estimates. It inherits nerf_trainer.py.
  • nerf_trainer_w_fixed_colmap_poses.py: The main engine for training a NeRF model, considering fixed poses, obtained from COLMAP (or some other initialization). The poses are therefore most likely noisy. Only the NeRF model is optimized. This is to show comparison to the joint refinement. It inherits nerf_trainer.py.
• train_settings: Contains config files, specifying the training of a network. For each experiment, we create a new config file in train_settings. The logging and saving will adapt accordingly so only the config file needs to be created in train_settings.
• admin: Includes functions for loading networks, tensorboard etc. and also contains environment settings.
• datasets: Contains integration of a number of datasets. All datasets inherent from the base Dataset class in datasets/base.py
• utils: Contains functions for logging, visualizations, geometric operations and colmap initalization.
  • geometry: Utils for geometric transformations
  • colmap_initialization: Utils to initializes poses with COLMAP, run with different matchers
• models: Contains NeRF model definition, different pose parametrization and the correspondence network wrapper.

Some tips on using and understanding the codebase:

• Throughout the code base, the pose convention is [right, down, forward]. This is the standard form of camera extrinsic projection matrices in the notation of multi-view geometry (opencv, colmap). It transforms a world-frame 3D point x to the camera frame and projects it to pixel coordinates u with the standard equation u = K (R^{w2c} x + t^{w2c}).

• We create a config file for each experiments (including changing each hyper-parameters). If you want to create a different experiment, create a new config file with a different name.

• The configuration will be loaded as the variable opt and self.settings throughout the codebase.

• The data for the whole scene is loaded at the beginning of the training, after creating the dataset. Throughout the codebase, it is stored in data_dict. Important keys are 'image' (B, 3, H, W), 'pose' (w2c ground-truth poses in format (B, 3, 4)) and 'intr' (B, 3, 3). When the poses are optimized, the current estimate of the poses are stored in data_dict.poses_w2c.

• The output of the NeRF model (computation graph for forward/backprop) is stored in output_dict throughout the codebase.

• The loss factory can be found in training/core/loss_factory.py. In the config, the loss to use is defined as a string in opt.loss_type. The final loss can be a combination of multiple loss functions. The losses values are stored in a dict loss_dict. The weights of each are defined in opt.loss_weight.<name>. The final overall loss is automatically computed based on the defined opt.loss_type and opt.loss_weight. To add a new loss function, create a new class with a main function compute_loss(). Add this new class to the loss factory. The output of the compute_loss() function should be loss_dict, stats_dict and plotting_dict. The actual loss value for backward is stored in loss_dict.<name>. You also need to add the new loss weight to opt.loss_weight.<name>. It will automatically be added to the overall loss and logged to Tensorboard.

• If you are using a multi-GPU machine, you can add --gpu=<gpu_number> to specify which GPU to use. Multi-GPU training/evaluation is currently not supported.

• The code was designed for a limited number of views. If loading many views instead and computing the correspondences exhaustively between the views, you might run into out-of-memory issues.

• If restarting the training of an experiment, it will automatically load the last saved checkpoint.

• (to be continued....)

---

V. License

This code is licensed under the Apache 2.0 License. See LICENSE for more details.