Background
In 2022, the Federal Emergency Management Agency (FEMA) responded to over 100 officially declared disasters within the U.S. and it's territories. One of the most time-sensitive post-disaster activities is the Preliminary Damage Assessment, to determine the level and extent of damage before restoration activities can begin. In theory, if some or all of this process could be automated, the faster PDA's are performed, the quicker individuals can obtain insurance reimbursement or disaster management officials are able to accurately prioritize restoration activities.
Project Description
This project is a proof-of-principle demonstration of whether the resnet-18 model for image recognition shows promise to differentiate between smoke, fire, damage, flooded and undamaged structures.
Instructions
cd ./jetson-inference/python/training/classification
wget https://www.dropbox.com/s/5d03a2n4klgkmfx/DamageSurvey_Data.tar?dl=0 -O DamageSurvey.tar
tar -xvf DamageSurvey.tar
python3 train.py --model-dir=models/DamageSurvey data/DamageSurvey
python3 onnx_export.py --model-dir=models/DamageSurvey
rm -r data/DamageSurvey/test/output
mkdir data/DamageSurvey/test/output
imagenet --model=models/DamageSurvey/resnet18.onnx --input_blob=input_0 --output_blob=output_0 --labels=data/DamageSurvey/labels.txt data/DamageSurvey/test/001.jpg data/DamageSurvey/test/output/output_001.jpg
imagenet --model=models/DamageSurvey/resnet18.onnx --input_blob=input_0 --output_blob=output_0 --labels=data/DamageSurvey/labels.txt data/DamageSurvey/test data/DamageSurvey/test/output
Downloads
@NVIDIA/Training Thanks to the Jetson Nano training and support teams for putting the material together to jumpstart projects!