This project is to apply semantic segementation on plant images. Original plant images are downloaded from google image search and cropped to a squared fix size. Mask images are generated using PlantCV package and manual manipulation in MS Paint. The backgroud is painted by black color while the plant area is highlighted by white color. plantcv is not needed to run trainer and predictor.

Deeplabv3 pre-trained deeplabv3_resnet50 model developed by Pytorch Team is tuned to fit this plant image dataset. Only 50 paired plant images as inputs can already enable the model to have good performance.

Mandatory python packages:

• torch (cpu/cuda)
• torchvision
• PIL
• matplotlib

Save your image to replace the file ./images/test/original.jpg with the same name. Then run the predictor, predicted mask and masked image will be updated for your custom image in the test image folder. Now, go try your photo!

Training Loss And Performance Tracking

Testing Trained Model With New Images