This repository contains our dataset contribution and codebase to reproduce our paper titled "Urban Tree Generator: Spatio-Temporal and Generative Deep Learning for Urban Tree Localization and Modeling" in CGI 2022 (published in The Visual Computer journal).

Read our paper here.

Besides using the codebase to reproduce our results, we hope that the dataset and codebase will help other researchers extend our methods in other domains also.

As per our dataset contribution in our paper noted in Sec. 3.1, the annotated dataset of four cities (Chicago, Indianapolis, Austin, and Lagos) into three classes - tree, grass, others can be downloaded from here.

A sample of the annotation of Indianapolis is shown below (green = tree, red = grass):

All the required libraries are enlisted in requirements.txt. To directly install using pip, please just use:

pip install -r requirements.txt

The repository is arranged so that can be easily reproducible into directories. The directory Segmentation_and_clustering contains all the code necessary to train and infer the segmentation and clustering section as noted in the paper. Here are some points as pre-requisites:

• Clone into this directory.
• Download the preprocessed training data from here.
• Place the zip file inside the Segmentation_and_clustering directory and unzip
• A directory called Data will be created
• Simply run ./Segmentation_and_clustering/python main.py to train
• Inference and usage of pre-trained models are documented and commented inside main.py

The directory Localization contains all the code necessary to train and infer the localization section as noted in the paper (Sec. 4). Here are some points as pre-requisites:

• Clone into this directory.
• Download the preprocessed training data from here.
• Place the zip file inside the Localization directory and unzip
• Simply run ./Localization/python train_localization.py to train the cGAN model
• Inference and usage of pre-trained models are documented and commented inside train_localization.py.

Below is our deep learning model for Segmentation of trees (Sec 3).

For the model of our localization network (Sec. 4) please see the implementatation inside train_localization.py which is inspired by the standard Tensorflow cGAN network. The figure is reproduced below.

Ref. Tensorflow cGAN network.

An illustrative example of our Localization output (bottom row) and ground truth (top row) is shown below in a segment of Chicago (see Sec. 4 and Sec. 5 of our paper for more results and examples).

If our paper, data and/or approach is in any way helpful to you and your research, please cite the paper from here as BibTeX.

Alternatively, cite as:

Firoze, A., Benes, B. & Aliaga, D. Urban tree generator: spatio-temporal and generative deep learning for urban tree localization and modeling. Vis Comput (2022). https://doi.org/10.1007/s00371-022-02526-x