This repository contains the code that accompanies our paper ATISS: Autoregressive Transformers for Indoor Scene Synthesis.
You can find detailed usage instructions for training your own models, using our pretrained models as well as performing the interactive tasks described in the paper below.
If you found this work influential or helpful for your research, please consider citing
@Inproceedings{Paschalidou2021NEURIPS,
author = {Despoina Paschalidou and Amlan Kar and Maria Shugrina and Karsten Kreis and Andreas Geiger and Sanja Fidler},
title = {ATISS: Autoregressive Transformers for Indoor Scene Synthesis},
booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
year = {2021}
}
Our codebase has the following dependencies:
For the visualizations, we use simple-3dviz, which is our easy-to-use library for visualizing 3D data using Python and ModernGL and matplotlib for the colormaps. Note that simple-3dviz provides a lightweight and easy-to-use scene viewer using wxpython. If you wish you use our scripts for visualizing the generated scenes, you will need to also install wxpython. Note that for all the renderings in the paper we used NVIDIA's OMNIVERSE.
The simplest way to make sure that you have all dependencies in place is to use
conda. You can
create a conda environment called atiss using
conda env create -f environment.yaml
conda activate atiss
Next compile the extension modules. You can do this via
python setup.py build_ext --inplace
pip install -e .
To evaluate a pretrained model or train a new model from scratch, you need to
obtain the
3D-FRONT
and the
3D-FUTURE
dataset. To download both datasets, please refer to the instructions provided in the dataset's
webpage.
As soon as you have downloaded the 3D-FRONT and the 3D-FUTURE dataset, you are
ready to start the preprocessing. In addition to a preprocessing script
(preprocess_data.py), we also provide a very useful script for visualising
3D-FRONT scenes (render_threedfront_scene.py), which you can easily execute by running
python render_threedfront_scene.py SCENE_ID path_to_output_dir path_to_3d_front_dataset_dir path_to_3d_future_dataset_dir path_to_3d_future_model_info path_to_floor_plan_texture_images
You can also visualize the walls, the windows as well as objects with textures
by setting the corresponding arguments. Apart from only visualizing the scene
with scene id SCENE_ID, the render_threedfront_scene.py script also
generates a subfolder in the output folder, specified via the
path_to_output_dir argument that contains the .obj files as well as the textures of all objects in this scene.
Once you have downloaded the 3D-FRONT and 3D-FUTURE datasets you need to run
the preprocess_data.py script in order to prepare the data to
be able to train your own models or generate new scenes using previously
trained models. To run the preprocessing script simply run
python preprocess_data.py path_to_output_dir path_to_3d_front_dataset_dir path_to_3d_future_dataset_dir path_to_3d_future_model_info path_to_floor_plan_texture_images --dataset_filtering threed_front_bedroom
Note that you can choose the filtering for the different room types (e.g.
bedrooms, living rooms, dining rooms, libraries) via the dataset_filtering
argument. The path_to_floor_plan_texture_images is the path to a folder
containing different floor plan textures that are necessary to render the rooms
using a top-down orthographic projection. An example of such a folder can be
found in the demo\floor_plan_texture_images folder.
This script starts by parsing all scenes from the 3D-FRONT dataset and then for
each scene it generates a subfolder inside the path_to_output_dir that
contains the information for all objects in the scene (boxes.npz), the room
mask (room_mask.png) and the scene rendered using a top-down
orthographic_projection (rendered_scene_256.png). Note that for the case of
the living rooms and dining rooms you also need to change the size of the room
during rendering to 6.2m from 3.1m, which is the default value, via the
--room_side argument.
Morover, you will notice that the preprocess_data.py script takes a
significant amount of time to parse all 3D-FRONT scenes. To reduce the waiting
time, we cache the parsed scenes and save them to the /tmp/threed_front.pkl
file. Therefore, once you parse the 3D-FRONT scenes once you can provide this
path in the environment variable PATH_TO_SCENES for the next time you run this script as follows:
PATH_TO_SCENES="/tmp/threed_front.pkl" python preprocess_data.py path_to_output_dir path_to_3d_front_dataset_dir path_to_3d_future_dataset_dir path_to_3d_future_model_info path_to_floor_plan_texture_images --dataset_filtering threed_front_bedroom
Finally, to further reduce the pre-processing time, note that it is possible to run this script in multiple threads, as it automatically checks whether a scene has been preprocessed and if it is it moves forward to the next scene.
As soon as you have installed all dependencies and have generated the
preprocessed data, you can now start training new models from scratch, evaluate
our pre-trained models and visualize the generated scenes using one of our
pre-trained models. All scripts expect a path to a config file. In the config
folder you can find the configuration files for the different room types. Make
sure to change the dataset_directory argument to the path where you saved the
preprocessed data from before.
To generate rooms using a previously trained model, we provide the
generate_scenes.py script and you can execute it by running
python generate_scenes.py path_to_config_yaml path_to_output_dir path_to_3d_future_pickled_data path_to_floor_plan_texture_images --weight_file path_to_weight_file
where the argument --weight_file specifies the path to a trained model and
the argument path_to_config_yaml defines the path to the config file used
to train that particular model. By default this script randomly selects floor
plans from the test set and conditioned on this floor plan it generate
different arrangements of objects. Note that if you want to generate a scene
conditioned on a specific floor plan, you can select it by providing its
scene id via the --scene_id argument. In case you want to run this script
headlessly you should set the --without_screen argument. Finally, the
path_to_3d_future_pickled_data specifies the path that contains the parsed
ThreedFutureDataset after being pickled.
To perform scene completion, we provide the scene_completion.py script that
can be executed by running
python scene_completion.py path_to_config_yaml path_to_output_dir path_to_3d_future_pickled_data path_to_floor_plan_texture_images --weight_file path_to_weight_file
where the argument --weight_file specifies the path to a trained model and
the argument path_to_config_yaml defines the path to the config file used
to train that particular model. For this script make sure that the encoding type in the config file has also the word eval in it.
By default this script randomly selects a room from the test set and
conditioned on this partial scene it populates the empty space with objects.
However, you can choose a specific room via the --scene_id argument.
This script can be also used to perform object placement. Namely starting from
a partial scene add an object of a specific object category.
In the output directory, the scene_completion.py script generates two folders
for each completion, one that contains the mesh files of the initial partial
scene and another one that contains the mesh files of the completed scene.
We also provide a script that performs object suggestions based on a user-specified region of acceptable positions. Similar to the previous scripts you can execute by running
python object_suggestion.py path_to_config_yaml path_to_output_dir path_to_3d_future_pickled_data path_to_floor_plan_texture_images --weight_file path_to_weight_file
where the argument --weight_file specifies the path to a trained model and
the argument path_to_config_yaml defines the path to the config file used
to train that particular model. Also for this script, please make sure that the
encoding type in the config file has also the word eval in it. By default
this script randomly selects a room from the test set and the user can either
choose to remove some objects or keep it unchanged. Subsequently, the user
needs to specify the acceptable positions to place an object using 6 comma
seperated numbers that define the bounding box of the valid positions.
Similar to the previous scripts, it is possible to select a particular scene by
choosing specific room via the --scene_id argument.
In the output directory, the object_suggestion.py script generates two folders
in each run, one that contains the mesh files of the initial
scene and another one that contains the mesh files of the completed scene with
the suggested object.
We also provide a script that performs failure cases correction on a scene that contains a problematic object. You can simply execute it by running
python failure_correction.py path_to_config_yaml path_to_output_dir path_to_3d_future_pickled_data path_to_floor_plan_texture_images --weight_file path_to_weight_file
where the argument --weight_file specifies the path to a trained model and
the argument path_to_config_yaml defines the path to the config file used
to train that particular model. Also for this script, please make sure that the
encoding type in the config file has also the word eval in it. By default
this script randomly selects a room from the test set and the user needs to
select an object inside the room that will be located in an unnatural position.
Given the scene with the unnatural position, our model identifies the
problematic object and repositions it in a more plausible position.
In the output directory, the falure_correction.py script generates two folders
in each run, one that contains the mesh files of the initial
scene with the problematic object and another one that contains the mesh files
of the new scene.
Finally, to train a new network from scratch, we provide the
train_network.py script. To execute this script, you need to specify the
path to the configuration file you wish to use and the path to the output
directory, where the trained models and the training statistics will be saved.
Namely, to train a new model from scratch, you simply need to run
python train_network.py path_to_config_yaml path_to_output_dir
Note that it is also possible to start from a previously trained model by
specifying the --weight_file argument, which should contain the path to a
previously trained model.
Note that, if you want to use the RAdam optimizer during training, you will have to also install to download and install the corresponding code from this repository.
We also provide the option to log the experiment's evolution using Weights &
Biases. To do that, you simply need to set the
--with_wandb_logger argument and of course to have installed
wandb in your conda environment.
Please also check out the following papers that explore similar ideas: