Bridging the Catalog to Real Gap Scalable Product Recognition via Multi-Stage Contrastive Learning

Jul 16, 2026

Upload by Maeve

Automated product recognition underpins retail intelligence, yet matching noisy in-store images to massive corporate catalogs is a scalability bottleneck. This work reframes recognition as embedding-based cross-domain retrieval: given a real-world product query crop, retrieve the most corresponding catalog reference image. The Cat2Real multi-stage contrastive learning framework fine-tunes a vision backbone using item-level and image-level similarities with targeted hard negative mining to bridge studio packshots and in-store queries. Evaluations show scalable zero-shot generalization to unseen products and categories, including settings with no real-world training images for new inventory.

LanguageEnglish

CategoryTechnology

FormatPDF

Pages12

Reading Time30 min

Introduction

Related Work

Product Recognition in Retail Industry

Bridging the Catalog-to-Real Gap: Scalable Product Recognition via Multi-Stage Contrastive Learning Anyi Zhang Walmart Global Tech Sunnyvale, CA, USA [anyi.zhang@walmart.com](anyi.zhang@walmart.com) Joy Mazumder Walmart Global Tech Sunnyvale, CA, USA [joy.mazumder@walmart.com](joy.mazumder@walmart.com) Kiril Lomakin Walmart Global Tech Sunnyvale, CA, USA [kiril.lomakin@walmart.com](kiril.lomakin@walmart.com) arXiv :2607 .09888v1 [ cs .CV] 10 Jul 2026 Abstract Automated product recognition is a cornerstone of modern retail intelligence; however, accurately matching realworld, in-store images against extensive corporate catalogs remains a major scalability bottleneck for large-scale applications. In this work, we address this challenge by reformulating the task as an embedding-based cross-domain retrieval problem rather than a standard closed-set classification task. Specifically, we define the objective as retrieving the most corresponding catalog reference image for a given real-world product query crop from an expansive inventory. To bridge the severe domain gap between pristine studio packshots and noisy in-store queries, we introduce a novel catalog-to-real multi-stage contrastive learning paradigm (Cat2Real). This framework fine-tunes a vision backbone by systematically exploiting both item-level and image-level similarities to drive targeted hard negative mining. Extensive empirical evaluations demonstrate that our paradigm scales seamlessly to unseen products and categories, yielding outstanding zero-shot generalization performance even in the complete absence of real-world training images for novel inventory. The fine-tuned model is made available under Apache 2.0 license 1. 1. Introduction Automated product recognition is a core catalyst for modern retail intelligence, enabling optimized supply chains and seamless customer experiences. While the initial pipeline step of localizing products is largely solved by modern object detectors, scaling the subsequent classification phase to global enterprise inventories remains a formidable bottleneck. This challenge stems from three compounding factors: (1) Extreme Scale and Volatility: Hypermarkets stock millions of active Stock Keeping Units (SKUs) that undergo daily inventory additions and frequent packaging 1[https://huggingface.co/zhanganyi88/Cat2Real-DINOv3-384](https://huggingface.co/zhanganyi88/Cat2Real-DINOv3-384) redesigns, rendering static closed-set classifiers obsolete. (2) Fine-Grained Ambiguity: Pervasive inter-class similarity exists across retail categories, where distinct items share near-identical profiles differing only by subtle visual cues, such as flavor, size, or minor textual details. (3) Environmental Noise: Real-world in-store images suffer from low resolution, severe occlusion, arbitrary rotations, motion blur, and glare from protective glass displays. To circumvent these challenges, a viable production system must operate under two core constraints: it must rely exclusively on pristine, pre-existing catalog packshots to eliminate the prohibitive costs of manual in-store annotation; and it must absorb real-time inventory updates without requiring model retraining. Consequently, we reformulate product recognition from a traditional closed-set classification problem into an open-set, cross-domain retrieval task. The objective is to project both noisy real-world product crops and high-quality catalog reference images into a shared latent embedding space where matching is performed via metric distance (see Fig. 1) . Because off-the-shelf foundation vision models lack sufficient discriminatory power across this severe domain gap, we introduce a novel multi-stage Catalog-to-Real (Cat2Real) contrastive learning framework. Unlike standard contrastive paradigms that rely on random negative sampling, which often yields uninformative or contradictory gradients in fine-grained tasks, Cat2Real systematically optimizes a vision ba

FAQs

What problem does the paper address in large-scale product recognition?

Accurately matching noisy real in-store product images to extensive enterprise catalog images at very large scale is difficult and becomes a major bottleneck for production systems.

How does Cat2Real reformulate product recognition compared with traditional classification?

It converts closed-set classification into an open-set, cross-domain retrieval task by embedding both query crops and catalog reference images into a shared latent space and performing matching by metric distance.

What key method does Cat2Real use to handle the studio-to-real domain gap?

It introduces a multi-stage contrastive learning framework that leverages item-level and image-level similarities, employs dynamic nearest-neighbor selection, and performs targeted hard negative mining to reduce training noise and improve invariance.

Jul 16, 2026

Upload by Maeve