link: https://arxiv.org/pdf/2305.03706

github: https://github.com/ladwigd/Leaflet-Product-Classification/tree/main?tab=readme-ov-file

What is the paper about?

A new dataset containing snippets from product leaflets has been created. Basic image and text-based modelling has been done to predict the product-dependent class/category from each individual product information snippet. The ‘snippets’ look like each of the product information containers in the below image:

Dataset

41.6k product images in 832 classes obtained from leaflets.

Base leaflets had come from 132 different retailers from 2016 to 2022

Classes include items including food, beverages, household goods, cosmetics, pet foods, etc.

Each class has 40 images in training and 10 in test

No parsed text from images has been included

Why is this dataset interesting?

• It’s novel in terms of the problem it is trying to solve.
• Same products will look different across different seller leaflets.
• Text and price information is there for every product in the snippet but it is noisy and tough to easily parse. Different colors, sizes, fonts, formatting, etc. are applied and there is no standard for the same.
• Discount, etc. information sometimes overlaps with the product photo and other information making
• Similar looking products can be standing for extremely different looking weight categories: 500 vs 750ml coke. Get tough to predict the product with the ml included.
• Weight and other text information can be represented in a granualar/fragmented manner: 290g could be written as 250 + 40g.

Baseline Models