• These models were trained for the kaggle compitition and won a BRONZE medal.

• Given a fragment, this segmentation model detects the ink.

• The output of the model is a binary mask where the 1 represents the presence of ink and 0 for no ink.

• This above mask is then converted into the RLE(Run Length Encodings).

  

The following example has the letter "E" it's RLE encoding is: "11 7 20 1 23 1 26 1 29 1 32 1 35 1 38 1 44 1"

• The data is taken from the kaggle compitition.

• The dataset is 3d x-ray scans of detached fragments of ancient papyrus scrolls.

• The training data had 3 fragments.

• It had the slices from the 3d x-ray surface volume. Each file contains a greyscale slice in the z-direction. Each fragment contains 65 slices. Combined this image stack gives us width * height * 65 number of voxels per fragment.

  

The 65 slices of the first fragment

• The inklables were given which was a binary mask which showed 1 for the presense of the ink.

  

• Further there was even the mask of the fragment which basically shows where the data is present in the fragment.

  

• Training_notebooks: Contains the notebooks used for training. Much more description about them is given in the training readme.
• Inference_notebooks: Contains the notebooks used for Inference. Much more description about them is given in the Inference readme.
• tiles_extracted: This is the notebook that reduces the dataset by merging the volume scans with the mask. This removes the unecessary tiles for training the model

• Trained multiple U-Net models using the segmentation-models-pytorch.

• Trained models with the backbone

  1. mit_b2 , mitb3 , mit_b4 , mit_b5
  2. VGG19
  3. resnet50 resnet34
  4. efficientnet_b3
  5. resnest
  6. regnety etc

• Few of them can be found in the repo.

• Trained a model using stratified K fold technique

• During the training the fragments were broken into the sizes of 224 by 224 and kept a stride of 224//4

• trained for 15 epochs by keeping one of the 3 fragments as the cross validation

• The calculated metric was fbeta score keeping the beta value 0.5

• Was getting the CV score of single model around 63

• used the weighted loss of dice loss and the Tversky loss and BCE loss

fig: the first image is the actual mask, the second one is the predicted mask, third one is the predicted mask after applying a threshhold 0.4

• Did the ensembles of different models trained, tried average ensemble.
• Ensembling the models increased the fbeta score from 65 to 74.
• Tried weighted ensembles tooo

• Performed TTA(Test Time Augmentations) during the final inference.

• So took the 8 channels and then predicted the output for the 1-3 , 4-6 , 6-8 channels and took the average of the predictions