SynthRetailProduct3D (SyRePro3D): A Pipeline for Synthesis of 3D Retail Product Models with Domain Specific Details Based on Package Class Templates

This research paper presents a novel pipeline for synthesizing high-quality, textured 3D models of retail products, addressing the limitations of existing datasets in the retail automation domain. As retail automation continues to revolutionize various aspects of business operations, there is a growing need for comprehensive and domain-specific datasets to enable the development and evaluation of robust perception systems. However, current datasets lack the necessary variety and size, often containing only a fraction of the products offered by retail companies and failing to capture domain-specific characteristics. To overcome these limitations, we introduce a modular pipeline that focuses on the synthesis of 3D models with high-resolution textures. Unlike previous methods that prioritize visually appealing shapes, our pipeline explicitly incorporates retail domain-specific details, including readable text, logos, codes, and nutrition tables, in the texture of the models. By disentangling the texturing process from shape augmentation, we ensure the visual quality of text areas, providing realistic and readable representations of retail products. In our methodology, we leverage state-of-the-art generative image models to add retail domain-specific details to the 3D models. Our approach enables the synthesis of diverse retail product packages with accurate textures, enhancing the realism and applicability of the generated models for the development and evaluation of perception systems in retail automation and robotics. Experimental results demonstrate the effectiveness and visual quality of models generated by our pipeline.