Hello. It was an interesting competition. I took the most interesting 11th place. Given the story with the alleged disqualifications, I am waiting for confirmation of the results. The 11th place in this situation is a borderline place) It seems that this test will never end, and this is not good from the point of view of lost time.

Based on this, I decided to publish the way I used to create masks during the competition. This method helped me take 11th place.

1. The first step is to install a cool ocr library called Doctr. This library reads the image and outputs bounding boxes for almost all text. Using keywords, fill in white as much text as possible, capture some of the tables and frames.

2. Extract the contours of the sections from the cleaned images using a combination of methods of 2 libraries: open-cv and skimage.

3. A total of three extraction options are used, and the result of each of them is saved in a separate directory.

4. The created masks are viewed. The most successful specimens are selected from the general pile. Image masks are then created using simple open-cv techniques of the selected masks.

5. A segmentation model from the Pytorch segmentation models library is trained.

The code for extracting masks can be seen on kaggle

https://www.kaggle.com/code/zaakciiru/barbados-survey-plans-make-masks