[WIP] Adding Differentiable Binarization

[ariG23498]

This pull request introduces a preliminary implementation of the paper available at arXiv:1911.08947. The primary focus has been on setting up the dataset, model, and loss functions based on the methodologies outlined in the paper. We are utilizing the ICDAR 2015 dataset for this purpose.

Key Points:

• Dataset Integration: The ICDAR 2015 dataset has been integrated for training and testing purposes.
• Model and Losses Setup: Both the model architecture and the loss functions have been defined in accordance with the paper.
• Current Progress:
  • The implementation is still in its early stages.
  • End-to-end training is not yet established.

Challenges and Immediate Focus:

• Data Batching: Currently, the data is not batched, which is a significant aspect we need to address for effective training.
• Loss Function Testing: The loss functions have not been tested thoroughly. This is a critical next step to ensure the model learns as expected.
• Single Data Point Testing: For immediate testing, we aim to pass a single data point through the model to observe its behavior and output.

Proposers:

• Aakash Kumar Nain
• Aritra Roy Gosthipaty
• Ritwik Raha