System setup

xView2 inference requires a tremendous amount of computing power. Currently, CPU inference is wildly impractical. To that end, unless you have a dedicated workstation with ample GPU power such as an Nvidia DGX station, we recommend a cloud based solution such as AWS or Google Cloud Compute utilizing a GPU optimized instance. Prices vary on instance type and area to be inferred. Example instances:

1. AWS EC2
   1. P4d.24xlarge
   2. P3.16xlarge
2. G Cloud
   1. Todo!

Installation

Install from source

Note: Only tested on Linux systems.

1. Close repository: git clone https://github.com/fdny-imt/xView2_FDNY.git.
2. Create Conda environment: conda create --name xv2 --file spec-file.txt.
3. Activate conda environment: conda activate xv2.

Docker

Todo.

Usage

Argument	Required	Default	Help
--pre_directory	Yes	None	Directory containing pre-disaster imagery. This is searched recursively.
--post_directory	Yes	None	Directory containing post-disaster imagery. This is searched recursively.
--output_directory	Yes	None	Directory to store output files. This will be created if it does not exist. Existing files may be overwritten.
--n_procs	Yes	8	Number of processors for multiprocessing
--batch_size	Yes	2	Number of chips to run inference on at once
--num_workers	Yes	4	Number of workers loading data into RAM. Recommend 4 * num_gpu
--pre_crs	No	None	The Coordinate Reference System (CRS) for the pre-disaster imagery. This will only be utilized if images lack CRS data.
--post_crs	No	None	The Coordinate Reference System (CRS) for the post-disaster imagery. This will only be utilized if images lack CRS data.
--destination_crs	No	EPSG:4326	The Coordinate Reference System (CRS) for the output overlays.
--output_resolution	No	None	Override minimum resolution calculator. This should be a lower resolution (higher number) than source imagery for decreased inference time. Must be in units of destinationCRS.
--dp_mode	No	False	Run models serially, but using DataParallel
--save_intermediates	No	False	Store intermediate runfiles
--aoi_file	No	None	Shapefile or GeoJSON file of AOI polygons

Example invocation for damage assessment

On 2 GPUs: CUDA_VISIBLE_DEVICES=0,1 python handler.py --pre_directory <pre dir> --post_directory <post dir> --output_directory <output dir> --aoi_file <aoi file (GeoJSON or shapefile)> --n_procs <n_proc> --batch_size 2 --num_workers 6

Notes:

• CRS between input types (pre/post/building footprints/AOI) need not match. However CRS within input types must match.

Sources

xView2 1st place solution

• Model weights from 1st place solution for "xView2: Assess Building Damage" challenge. https://github.com/DIUx-xView/xView2_first_place
• More information from original submission see commit: 3fe4a7327f1a19b8c516e0b0930c38c29ac3662b