midv-500-models's Introduction

midv-500-models

The repository contains a model for binary semantic segmentation of the documents.

Left: input.

Center: prediction.

Right: overlay of the image and predicted mask.

Dataset

Model is trained on MIDV-500: A Dataset for Identity Documents Analysis and Recognition on Mobile Devices in Video Stream.

Preparation

Download the dataset from the ftp server with

wget -r ftp://smartengines.com/midv-500/

Unpack the dataset

cd smartengines.com/midv-500/dataset/
unzip \*.zip

The resulting folder structure will be

smartengines.com
    midv-500
        dataset
            01_alb_id
                ground_truth
                    CA
                        CA01_01.tif
                    ...
                images
                    CA
                        CA01_01.json
                    ...
                ...
            ...
        ...
    ...

To preprocess the data use the script

python midv500models/preprocess_data.py -i <input_folder> \
                                          -o <output_folder>

where input_folder corresponds to the file with the unpacked dataset and output folder will look as:

images
    CA01_01.jpg
    ...
masks
    CA01_01.png

target binary masks will have values [0, 255], where 0 is background and 255 is the document.

Training

python midv500models/train.py -c midv500models/configs/2020-05-19.yaml \
                              -i <path to train>

Inference

python midv500models/inference.py -c midv500models/configs/2020-05-19.yaml \
                                  -i <path to images> \
                                  -o <path to save preidctions>
                                  -w <path to weights>