price_detection_deep_learning

Data

The structure of the data folder is the following:

- data
    - train
        - images
            - 0000.jpg
            - 0001.jpg
        - price_tags
            - 0000.csv
            - 0001.csv
        - annocations.csv
        - price_tags.csv
    - test
        - images
            - 0000.jpg
            - 0001.jpg
        - results.csv

You can download the data from https://www.kaggle.com/itamargr/traxpricing-dataset, it is from the pricing challenge 2021 of Retail Vision for CVPR 2021 workshop https://retailvisionworkshop.github.io/pricing_challenge_2021/.

Dev

• Install the dependencies make install-dev
• Install the torch dependencies
  • On CPU: make install-cpu
  • On GPU: make install-gpu

Price annotation

To perform the price detection on the image, we need a dataset of the boxes on the images of the prices tickets. We are using jupyter_bbox_widget, a module that allows to annotate images directly in a jupyter notebook, and is easy to use + easy to retrieve annotations and put new annotations from our model. The idea is to proceed using the following steps:

• Annotate a few images (boxes coordinates and size)
• Train a model to predict the boxes positions
• Check the predictions and add them to the dataset
• Loop back until the model is good enough