Deep Learning-Based Feature Extraction for Accurate Estimation of Product Dimensions from Images

In the E-commerce sector, product dimensions play a crucial role both for inventory management and for enabling customers to filter products according to their requirements. Hence, accurately extracting dimensions from product data can help efficiently manage inventory space, save packaging costs and enhance customer experience. Existing methods use textual data like product title, description, and additional information for extracting dimension attributes. However, this textual information is usually ambiguous and disorganized. A few approaches also use images for dimension extraction, however these approaches only tackle length-based measurements, and do not provide an approach for handling other dimensions such as item weight, volume, and voltage. Hence, we propose a deep learning based approach to extract dimensions from product images, which uses Region Proposal Networks (RPNs) to create bounding boxes around products in images and Optical Character Recognition (OCR) pipelines with Regular Expressions (RegEx) to extract text bounding boxes. We then use relative positional information to identify the dimension in cases of length-based measurements. We showcase the results of our approach applied to a dataset of actual product images, and compare its performance with existing approaches.