包装水果和蔬菜识别基准

1. 简介该数据集是专注于水果和蔬菜的零售自动化视觉识别数据集，旨在克服现有开放数据集局限，支持线下零售环境多任务模型开发；包含 34 个物种、65 个品种（类别分布相对均衡），涵盖塑料袋包装商品，共 10 万 + 张图像（由 72 名标注者收集）、37 万个物体，覆盖多个商店和城市；每个样本从多视角捕捉，额外标注物体数量、总重量等信息，约 9000 张图像含手动分割遮罩；提供零样本分类、监督分类、实例分割、物体计数任务的基线结果（65 个品种零样本最佳 28%、监督最佳 96%），并研究了包装和背景类型对模型性能的影响1. Introduction This dataset is a visual recognition dataset dedicated to retail automation for fruits and vegetables, aiming to address the limitations of current open datasets and support the development of multi-task models in offline retail environments. It covers 34 species and 65 varieties with relatively balanced category distribution, including plastic bag-packaged commodities. The dataset contains over 100,000 images (collected by 72 annotators) and 370,000 object instances, spanning multiple stores and cities. Each sample is captured from multiple perspectives, with additional annotations such as object quantity and total weight; approximately 9,000 images come with manually segmented masks. Baseline results for four tasks are provided: zero-shot classification, supervised classification, instance segmentation, and object counting (the best zero-shot accuracy for the 65 varieties reaches 28%, while the best supervised accuracy hits 96%). Additionally, the impact of packaging and background types on model performance has been studied.

搜集汇总

背景与挑战

背景概述

该数据集是一个专注于水果和蔬菜零售自动化视觉识别的基准数据集，旨在支持线下零售环境的多任务模型开发。它包含34个物种、65个品种的10万+张图像，涵盖塑料袋包装商品，覆盖多商店和城市场景，并提供类别、属性及分割标注，支持零样本分类、监督分类、实例分割和物体计数等任务。