Retail Product Detection

Ref - Post 

• Product region, brand logo region
• Product textual data (title, brands)
• The regions of interest in images were detected by a pretrained teacher model
• Following the trend of using free-form text, we train the CPG model with 2.3M product entities synthesized from an e-commerce site in a self-supervised fashion
• The bounding boxes for product-noun-to-object task are generated by a pre-trained general domain modulated detection model
• Visual-language understanding of logos, brand strings, product details for the query product entity and for all brand representative product entities

• Text to image lookup and comparison
• Similar embedding lookup and comparison

• Crafted image caption is tokenized and encoded using a pre-trained text encoder: RoBERTa
• Image and textual features are concatenated as a multimodal vector and fed to a joint transformer encoder with cross attention between image and textual features

Keep Exploring!!!

Production Detection from logos

Ref from post, post1

• Detecting all the possible image regions that might contain a logo
• Matching the detected regions against an ever-evolving set of logo prototypes
• Representation of an image with its text description
• Zero-shot logo recognition relies heavily on an embedding model for matching query regions against a constantly evolving set of cropped logo images
• Text pipeline to extract the text in the image via optical character recognition (OCR)
• We explicitly pair positive examples with their hard negatives, to encourage the model to distinguish logos from similar texts
• Create embedding for similar products and assign appropriate weights

Keep Exploring!!!