Semi-Automatic Segmentation model testing

Test models and image pre-processing techniques to find the best combination.

Usage

1. Define the WORK_DIR with the working directory of your project. It should contain:
   1. File named dataset.zip containing:
      • input folder with all the input images
      • label folder with all the target images (masks) with the same name as in the input folder
      • test-input folder with the test input images
      • test-label folder with the target test masks with the same name as in the input folder
   2. Folder named models
2. Define SIZE with the model input size (SegNet follows 256)
3. Define STRIDE with the stride that the algorithm will make to generate the dataset
4. Define the CURRENT_MODEL with the path of your model according to image-segmentation-keras
   • CURRENT_MODEL = "segnet.resnet50_segnet" will produce a SegNet model using the ResNet50 backbone
5. Run all the cells until Functions stage
6. If your model is new (first time using it)
   1. Run create_default_model()
7. If your dataset is new (first time using it)
   1. Run develop_dataset() (last cell)
   2. Wait for it to complete, may take a while
8. Create/modificate the image pre-processing cells, starting at Standard
9. Run the cells
   • When a cell is completed, it should create a folder with the model results
10. At the end, run fetch_model_results() to get all the model results into its own folder

Analysing data

1. Copy the CURRENT_MODEL.csv file to the same dir as data_analysis.ipynb
2. Run the cells
3. Check the result.png file

Example

On the right there's the model trained on 1024 images and on the left there's the model results trained on 7000 images. Each line corresponds to a different image processing technique.