SBIR Phase I: Innovative visual search and similarity for decor, apparel, and style

这个小企业创新研究（SBIR）第一阶段项目的更广泛的影响/商业潜力是开发和商业化视觉搜索，以细粒度识别室内装饰和服装的产品和风格。这项技术将帮助更广泛的公众找到使用传统的基于文本的搜索可能难以搜索的项目。在许多市场（家居装饰、时尚等）中，客户寻求具有独特视觉外观的产品，这些产品无法通过基于文本的搜索轻松表达。该项目将为家居装饰和服装市场开发视觉搜索工具。这个小企业创新研究（SBIR）第一阶段项目将开发基于深度学习的软件，用于产品和服装识别以及风格识别。我们之前的原型使用深度学习从客户拍摄的“常规”照片中识别特定产品，其中的挑战是，与标志性产品图像相比，这些产品的常规照片可能有许多不同的背景，大小，方向或照明，并且产品可能会被场景中的杂乱明显遮挡。该项目的目标是通过四个主要目标概括工作，以实现广泛的适用性：概括设置和产品分类和分类法，以支持广泛的客户和产品类型（目标1）；半自动检测场景图像中的产品，以扩展到大型照片集（目标2）；细化训练模型以获得细粒度匹配以满足客户需求（目标3）；并将系统实时部署到公司（目标4）。

The broader impact/commercial potential of this Small Business Innovation Research (SBIR) Phase I project is to develop and commercialize visual search for fine-grained recognition of products and style in interior decor and apparel. The technology will help the broader public find items that may be difficult to search for using traditional text-based search. In many markets (home decor, fashion, etc.), customers seek products that have unique visual appearances that cannot easily be expressed with a text-based search. This project will develop visual search tools for the home decor and apparel markets. This Small Business Innovation Research (SBIR) Phase I project will develop software based on deep learning for product and apparel recognition and style recognition. Our prior prototype uses deep learning to recognize specific products from "regular" photographs taken by customers, where the challenge is that these regular photos of products can have many different backgrounds, sizes, orientations, or lighting when compared to the iconic product image, and the product could be significantly occluded by clutter in the scene. The goal of this project is to generalize the work to achieve broad applicability through four major objectives: Generalizing the settings and product categorization and taxonomy to support a broad range of customers and product types (Objective 1); semi-automatic detection of products in scene images to scale to large photo collections (Objective 2); refining the trained models for fine-grained matches to meet customer needs (Objective 3); and deploying the system live to companies (Objective 4).