This code is for TMD project

Please follow this link for the installation https://www.highvoltagecode.com/post/edge-ai-semantic-segmentation-on-nvidia-jetson Note: if there is any problem of the cuda, please switch to CPU

There are 3 classes for labelMe. class.txt Use LabelMe to create the 3 segmentation of each picture

labelme --labels classes.txt

Convert to VOC format

python labelme2voc.py {data folder} --labels classes.txt --noviz

Split the data based on 10 blocks, and create 10 trial folder dir0-dir9

python split_custom.py --masks="{data folder}/SegmentationClass" --images="{data folder}/JPEGImages" --output="{data folder}/dir" --keep-original

For training:

python train.py {data folder}/dir0 --dataset=custom --epochs=100 --classes=4 --arch=fcn_resnet101 --device=cpu
python onnx_export.py --output={output folder}/fcn_reset101_0.onnx

python train.py {data folder}/dir1 --dataset=custom --epochs=100 --classes=4 --arch=fcn_resnet101 --device=cpu
python onnx_export.py --output={output folder}/fcn_reset101_1.onnx

python train.py {data folder}/dir2 --dataset=custom --epochs=100 --classes=4 --arch=fcn_resnet101 --device=cpu
python onnx_export.py --output={output folder}/fcn_reset101_2.onnx

python train.py {data folder}/dir3 --dataset=custom --epochs=100 --classes=4 --arch=fcn_resnet101 --device=cpu
python onnx_export.py --output={output folder}/fcn_reset101_3.onnx

python train.py {data folder}/dir4 --dataset=custom --epochs=100 --classes=4 --arch=fcn_resnet101 --device=cpu
python onnx_export.py --output={output folder}/fcn_reset101_4.onnx

python train.py {data folder}/dir5 --dataset=custom --epochs=100 --classes=4 --arch=fcn_resnet101 --device=cpu
python onnx_export.py --output={output folder}/fcn_reset101_5.onnx

python train.py {data folder}/dir6 --dataset=custom --epochs=100 --classes=4 --arch=fcn_resnet101 --device=cpu
python onnx_export.py --output={output folder}/fcn_reset101_6.onnx

python train.py {data folder}/dir7 --dataset=custom --epochs=100 --classes=4 --arch=fcn_resnet101 --device=cpu
python onnx_export.py --output={output folder}/fcn_reset101_7.onnx

python train.py {data folder}/dir8 --dataset=custom --epochs=100 --classes=4 --arch=fcn_resnet101 --device=cpu
python onnx_export.py --output={output folder}/fcn_reset101_8.onnx

python train.py {data folder}/dir9 --dataset=custom --epochs=100 --classes=4 --arch=fcn_resnet101 --device=cpu
python onnx_export.py --output={output folder}/fcn_reset101_9.onnx

After the 10 onnx files are created. A Nvidia Nano is used for inference.

Build the Nvidia segmentation from source https://github.com/dusty-nv/jetson-inference/blob/master/docs/building-repo-2.md

For Inference, details, pls reference this link, https://github.com/dusty-nv/jetson-inference/blob/master/docs/segnet-console-2.md

python3 segnet_GY_dot.py {input folder} {output folder} --model={onnx file} --labels=classes.txt --colors=colors.txt 

For batch images

./batch_inference.sh {input_folder} {output_folder} {onnx file}