Deep Learning for Urban Tree Detection

This repository provides code for detection of trees in the urban environment from LIDAR and multispectral imagery. The network backbone is a modified PointNet and PointNet2 architecture, as developed by Qi et al. 2017.

Setting Up

Create a conda environment with tensorflow 2.4 (or 2.4-gpu), using the provided yml file

How to Run

Data Preprocessing

cd data_preprocessing

With an environment that has pdal and numpy, run the following:

python3 filter_lidar.py --infile <region las file> [--reproject ? --scale ?] With raw las/laz file, compute height above ground (HAG) dimension, filter to HAG>0, and possibly reproject (note the --scale argument when reprojecting).

Create and activate a laspy environment with the provided .yml file, then:

python3 generate_lidar_patches.py --specs ../data/elcap_lidar_specs.json --outname <dataset name> This selects all LIDAR points lying inside an annotated grid square, and saves it to a .npy file.

Optionally: python3 analyze_dataset.py --dsname <dataset name> to generate some nice diagnostic plots and such, inside data/lidar/<dataset name>

Training

First time only: ./docker_build.sh

Then, use ./docker_run N where N is the GPU index you would like to use to activate the docker container. If this is the first time you have ever activated the container, run ./compile_ops.sh inside as well. ./docker_run also accepts following position arguments as a command, so ./docker_run 3 ./compile_ops.sh runs compile_ops.sh inside the container.

Finally, run

python3 train.py -h to see the full array of command line options available for training

Evaluation

Evaluation is run automatically after training. You can also run it on its own, with:

python3 evaluate.py --name <name> which evaluates the model with the given name