EmptyCities for visual SLAM

Torch implementation for learning a mapping from input images that contain dynamic objects in a city environment, such as vehicles and pedestrians, to output images which are static and suitable for localization and mapping.

Empty Cities: a Dynamic-Object-Invariant Space for Visual SLAM
Berta Bescos, Cesar Cadena, Jose Neira

Setup

Prerequisites

• Linux or OSX
• NVIDIA GPU + CUDA CuDNN (CPU mode and CUDA without CuDNN may work with minimal modification, but untested)

Getting Started

• Install torch and dependencies from https://github.com/torch/distro
• Install torch packages nngraph and display

luarocks install nngraph
luarocks install https://raw.githubusercontent.com/szym/display/master/display-scm-0.rockspec

• Clone this repo:

git clone https://github.com/BertaBescos/EmptyCities_SLAM.git
cd EmptyCities_SLAM

Models

Pre-trained models can de downloaded from the folder checkpoints in this link. You will find a README.md file inside this folder. Place the checkpoints folder inside the project.

Test

• We encourage you to keep your data in a folder of your choice /path/to/data/ with three subfolders train, test and val. The following command will run our model within all the images inside the folder test and keep the results in ./results/. Images within the folder test should be RGB images of any size.

DATA_ROOT=/path/to/data/ name=my_name th test.lua

• If you prefer to feed the dynamic/static binary masks, you should concatenate it to the RGB image. We provide a python script for this on https://github.com/bertabescos/EmptyCities.

DATA_ROOT=/path/to/data/ name=my_name mask=1 th test.lua

Train

• The simplest case trains only with synthetic CARLA data with G(x,m) and D(x,y,m,n). In the subfolder /path/to/synth/data/train/ there should be the concatenated (RGB | GT | Mask) images. The utilized masks come from this simulator too, and therefore do not use the semantic segmentation model.

DATA_ROOT=/path/to/synth/data/ name=my_name th train.lua

• If you want to use the ORB-features-based loss you should set lossDetector, lossOrientation and lossDescriptor to 0 in the command line.

DATA_ROOT=/path/to/synth/data/ name=my_name lossDetector=1 lossOrientation=1 lossDescriptor=1 th train.lua

• If you want to finetune your trained model with real-world data in your training, you should set NSYNTH_DATA_ROOT to this dataset path.

DATA_ROOT=/path/to/synth/data/ NSYNTH_DATA_ROOT=/path/to/non/synth/data/ continue_train=1 name=my_name th train.lua

• (Optionally) start the display server to view results as the model trains. ( See Display UI for more details):

th -ldisplay.start 8000 0.0.0.0

Models are saved by default to ./checkpoints/base_512x512 (can be changed by passing checkpoint_dir=your_dir and name=your_name in options.lua).

See options.lua for additional training options.

Datasets

Our synthetic dataset has been generated with CARLA 0.8.2 and is available in the zipped folder CARLA_dataset in this link. Information on how this dataset has been generated can be found in here.

Citation

If you use this code for your research, please cite ours papers:

@article{bescos2019empty,
  title={Empty Cities: a Dynamic-Object-Invariant Space for Visual SLAM},
  author={Bescos, Berta  and Cadena, Cesar and Neira, José},
  journal={arXiv},
  year={2019}
}

@article{bescos2018empty,
  title={Empty Cities: Image Inpainting for a Dynamic-Object-Invariant Space},
  author={Bescos, Berta and Neira, José and Siegwart, Roland and Cadena, Cesar},
  journal={International Conference on Robotics and Automation (ICRA)},
  year={2018}
}

Acknowledgments

Our code is heavily inspired by pix2pix, DCGAN and Context-Encoder.

EmptyCities for visual SLAM