细粒度商品图像识别数据集Products-10K

面向电商场景海量商品精细粒度识别任务，本项目构建了细粒度商品图像识别数据集Products-10K，本数据集图像数据采集自京东商城的10大类产品下属的10000类常见商品SKU，涵盖商家商品展示图以及用户下单实拍图，总数据量约20GB，采集时间为2021年6月。相比于其他主流商品图像识别数据集，本数据完全人工标注，数据集噪声比例控制在 0.5% 以内，每张图片至少经过来自京东商品识别专家团队三人以上的审查，是截至目前最大规模的有标注细粒度商品图像数据集。

To tackle the fine-grained recognition task for massive commodities in e-commerce scenarios, this project develops the fine-grained commodity image recognition dataset Products-10K. The image data of this dataset is collected from 10,000 common commodity SKUs under 10 major product categories on JD.com, including both merchant-provided product display images and user-taken real-world product photos captured during order placement. The total data volume is approximately 20 GB, and the data was collected in June 2021. Compared with other mainstream commodity image recognition datasets, this dataset is fully manually annotated, with its noise ratio controlled within 0.5%. Each image has been reviewed by at least three experts from JD.com's commodity recognition expert team, making it the largest labeled fine-grained commodity image dataset to date.

搜集汇总

背景与挑战

背景概述

细粒度商品图像识别数据集Products-10K是一个面向电商场景的大规模商品图像数据集，包含来自京东商城的10大类10000类商品SKU，总数据量约20GB。该数据集完全人工标注，噪声比例低于0.5%，是当前最大规模的有标注细粒度商品图像数据集。