mikelmh025/ClothingADC

--- license: cc-by-4.0 pretty_name: Clothing-ADC task_categories: - image-classification - image-feature-extraction language: - en source_datasets: - original size_categories: - 1M<n<10M paperswithcode_id: clothing-adc arxiv: 2408.11338 dataset_info: features: - name: id dtype: string - name: image dtype: image - name: class dtype: string - name: color dtype: string - name: material dtype: string - name: pattern dtype: string splits: - name: train num_examples: 1036738 - name: validation num_examples: 20000 - name: test num_examples: 20000 tags: - clothing - fashion - computer-vision - fine-grained-recognition - label-noise - noisy-labels - long-tail - class-imbalance - benchmark - web-crawled - data-curation - robustness --- # Clothing-ADC Dataset **Paper:** [Automatic Dataset Construction (ADC): Sample Collection, Data Curation, and Beyond](https://arxiv.org/abs/2408.11338) **Authors:** Minghao Liu, Zonglin Di, Jiaheng Wei, Zhongruo Wang, Hengxiang Zhang, Ruixuan Xiao, Haoyu Wang, Jinlong Pang, Hao Chen, Ankit Shah, Hongxin Wei, Xinlei He, Zhaowei Zhu, Haobo Wang, Lei Feng, Jindong Wang, James Davis, Yang Liu **Institutions:** UC Santa Cruz, HKUST(GZ), UC Davis, SUSTech, Zhejiang University, Yale University, Carnegie Mellon University, Nanyang Technological University, Microsoft --- ## Dataset Summary **Clothing-ADC** is a large-scale clothing image classification dataset built with the **Automatic Dataset Construction (ADC)** pipeline. Instead of the traditional approach of collecting images first and then annotating them, ADC reverses this process: it uses GPT-4 to design fine-grained class hierarchies, then automatically collects labeled images from Google Image Search using the class descriptions as queries. The result is a dataset with **over 1 million images**, **12 main clothing classes**, and **12,000 fine-grained subclasses** defined by combinations of color, material, and pattern attributes — all without requiring domain expertise or manual annotation of individual samples. The dataset also serves as a benchmark platform for three real-world data challenges that arise during automatic dataset construction: 1. **Label Noise Detection** 2. **Learning with Noisy Labels** 3. **Class-Imbalanced Learning** --- ## Dataset Statistics | Property | Value | |---|---| | Total samples | 1,076,738 | | Image resolution | 256 × 256 | | Main classes | 12 | | Total subclasses | 12,000 | | Avg. samples per subclass | ~89.73 | | Label noise rate (train) | 22.2% – 32.7% | ### Dataset Splits | Split | Size | Label Quality | |---|---|---| | Train | 1,036,738 | Web-collected (noisy) | | Validation | 20,000 | Human-verified (clean) | | Test | 20,000 | Human-verified (clean) | ### Main Classes (12) Sweater, Windbreaker, T-shirt, Shirt, Knitwear, Hoodie, Jacket, Suit, Shawl, Dress, Vest, Underwear ### Subclass Structure Each main class has **1,000 subclasses** defined by three attributes with 10 options each: | Attribute | # Options | Examples | |---|---|---| | Color | 10 | white, black, red, navy, grey, … | | Material | 10 | cotton, wool, polyester, denim, … | | Pattern | 10 | solid, striped, plaid, floral, … | Search queries are formed as `"<Color> <Material> <Pattern> <Clothing Type>"` (e.g., `"white cotton fisherman sweater"`), which serve simultaneously as the image search query and the sample's fine-grained label. --- ## ADC Pipeline The ADC pipeline consists of three steps: **Step 1 — Dataset Design with LLMs** GPT-4 is prompted to enumerate attribute options for each clothing category (`"Show me <30–80> ways to describe <Attribute> of <Class>"`). The resulting categories are reviewed iteratively, avoiding the need for human domain expertise. **Step 2 — Automated Labeling** The Google Image API is queried with composite search strings. The top ~100 results per query are collected and labeled automatically. Each query string is the sample's label, eliminating manual annotation entirely. **Step 3 — Data Curation and Cleaning** - **Algorithmic curation:** Label noise detection methods (e.g., Simi-Feat / Docta) automatically filter mislabeled samples, reducing noise from ~22.2% to ~10.7%. - **Human-in-the-loop (clean splits):** For the validation and test sets, human annotators on Amazon MTurk verified labels by selecting correct samples from machine-labeled batches (minimum 4 of 20 per query). Only samples with full human–machine agreement are included in the clean splits. --- ## Benchmark Tasks ### 1. Label Noise Detection (`Clothing-ADC-Detection`) A 20,000-sample subset with both noisy and clean labels, annotated by 3 Amazon MTurk workers per image (correct / unsure / incorrect). Used to benchmark noise detection algorithms. **Metric:** F1-score of detected corrupted instances | Method | F1-Score | |---|---| | CORES | 0.4793 | | Confident Learning (CL) | 0.4352 | | Deep k-NN | 0.3991 | | Simi-Feat | **0.5721** | --- ### 2. Label Noise Learning (`Clothing-ADC` / `Clothing-ADC-Tiny`) Train on the full noisy training set; evaluate on the clean held-out test set. A tiny version (~50K train images) is also provided for fast experimentation. **Metric:** Classification accuracy on clean test set (12-class) Selected results (ResNet-50, 20 epochs): | Method | Full | Tiny | |---|---|---| | Cross-Entropy (baseline) | — | — | | Positive Label Smoothing | ↑ | ↑ | | Taylor CE | **best** | **best** | | DivideMix | competitive | competitive | --- ### 3. Class-Imbalanced Learning (`Clothing-ADC-CLT`) A class-level long-tail version of the dataset, with imbalance ratios ρ ∈ {10, 50, 100}. Noisy samples are removed prior to constructing this benchmark using algorithmic curation (Docta + learning-centric curation), yielding ~562,263 clean images. **Metric:** δ-worst accuracy (interpolates between mean accuracy at δ=0 and worst-class accuracy at δ→∞) | Method | ρ=10 (δ=0) | ρ=100 (δ=0) | ρ=10 (δ=∞) | ρ=100 (δ=∞) | |---|---|---|---|---| | Cross-Entropy | 57.80 | 30.10 | 0.96 | 0.00 | | Focal Loss | 72.70 | 62.28 | 38.12 | 13.44 | | LDAM | 72.50 | 63.25 | 40.90 | 15.69 | | Balanced Softmax | 74.18 | 69.47 | 48.54 | 50.60 | | Logit-Adjust | **74.08** | **69.44** | 47.45 | 43.26 | | Drops | 73.66 | 67.15 | **50.85** | 32.43 | --- ## Comparison with Existing Datasets | Dataset | # Train/Test | # Classes | Noise Rate (%) | Has Attributes | Auto Annotation | Requires Expert? | |---|---|---|---|---|---|---| | iNaturalist | 579k/279k | 54k | ~0 | ✗ | ✗ | ✓ | | WebVision | 2.4M/100k | 1000 | 20 | ✗ | ✓ | ✓ | | ANIMAL-10N | 50k/10k | 10 | 8 | ✗ | ✗ | ✗ | | CIFAR-10N | 50k/10k | 10 | 9–40 | ✗ | ✗ | ✗ | | Food-101N | 75.75k/25.25k | 101 | 18.4 | ✗ | ✗ | ✓ | | Clothing1M | 1M total | 14 | 38.5 | ✗ | ✗ | ✓ | | **Clothing-ADC (Ours)** | **1M/20k** | **12** | **22.2–32.7** | **12k** | **✓** | **✗** | --- ## Data Fields Each sample contains: - `id`: unique image identifier string - `image`: PIL image (256×256 RGB) - `class`: main clothing category string (e.g., `"Sweater"`) - `color`: color attribute label (e.g., `"white"`) - `material`: material attribute label (e.g., `"cotton"`) - `pattern`: pattern attribute label (e.g., `"fisherman"`) --- ## Usage ```python from datasets import load_dataset # Full dataset ds = load_dataset("mikelmh025/ClothingADC") # Access splits train = ds["train"] val = ds["validation"] test = ds["test"] # Example: iterate over test set for sample in test: image = sample["image"] category = sample["class"] color = sample["color"] material = sample["material"] pattern = sample["pattern"] ``` --- ## Citation If you use Clothing-ADC in your research, please cite: ```bibtex @article{liu2024adc, title = {Automatic Dataset Construction (ADC): Sample Collection, Data Curation, and Beyond}, author = {Minghao Liu and Zonglin Di and Jiaheng Wei and Zhongruo Wang and Hengxiang Zhang and Ruixuan Xiao and Haoyu Wang and Jinlong Pang and Hao Chen and Ankit Shah and Hongxin Wei and Xinlei He and Zhaowei Zhu and Haobo Wang and Lei Feng and Jindong Wang and James Davis and Yang Liu}, journal = {arXiv preprint arXiv:2408.11338}, year = {2024} } ``` --- ## License This dataset is released under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). Images are collected from Google Image Search and remain subject to their original source licenses. This dataset is intended for research purposes only.许可证：CC BY 4.0 美观名称：Clothing-ADC 任务类别： - 图像分类 - 图像特征提取 语言： - 英语 源数据集： - 原创数据集 样本规模区间： - 100万<n<1000万 paperswithcode 标识符：clothing-adc arxiv 编号：2408.11338 数据集信息： 特征字段： - 名称：id，数据类型：字符串 - 名称：image，数据类型：图像 - 名称：class，数据类型：字符串 - 名称：color，数据类型：字符串 - 名称：material，数据类型：字符串 - 名称：pattern，数据类型：字符串 划分集： - 名称：训练集，样本数：1036738 - 名称：验证集，样本数：20000 - 名称：测试集，样本数：20000 标签： - 服装 - 时尚 - 计算机视觉 - 细粒度识别 - 标签噪声 - 噪声标签 - 长尾分布 - 类别不平衡 - 基准测试 - 网络爬取 - 数据整理 - 鲁棒性 # Clothing-ADC 数据集 **论文：** [Automatic Dataset Construction (ADC): Sample Collection, Data Curation, and Beyond](https://arxiv.org/abs/2408.11338) **作者：** Minghao Liu, Zonglin Di, Jiaheng Wei, Zhongruo Wang, Hengxiang Zhang, Ruixuan Xiao, Haoyu Wang, Jinlong Pang, Hao Chen, Ankit Shah, Hongxin Wei, Xinlei He, Zhaowei Zhu, Haobo Wang, Lei Feng, Jindong Wang, James Davis, Yang Liu **机构：** 加州大学圣克鲁兹分校、香港科技大学（广州）、加州大学戴维斯分校、南方科技大学、浙江大学、耶鲁大学、卡内基梅隆大学、南洋理工大学、微软公司 --- ## 数据集概述 **Clothing-ADC**是一款基于**自动数据集构建（Automatic Dataset Construction，ADC）**流程构建的大规模服装图像分类数据集。与传统先收集图像再进行标注的流程不同，ADC将该过程反转：首先使用大语言模型（Large Language Model，LLM）设计细粒度的类别层级结构，随后以类别描述作为查询词，从谷歌图像搜索中自动采集带标注的图像。 最终得到的数据集拥有**超过100万张图像**、**12个主服装类别**，以及**12000个由颜色、材质和图案属性组合定义的细粒度子类**，全程无需依赖领域专业知识或对单样本进行手动标注。 该数据集同时可作为三个自动数据集构建过程中常见的真实世界数据挑战的基准测试平台： 1. **标签噪声检测** 2. **噪声标签下的学习** 3. **类别不平衡学习** --- ## 数据集统计 | 属性 | 数值 | |---|---| | 总样本数 | 1,076,738 | | 图像分辨率 | 256 × 256 | | 主类别数 | 12 | | 总子类数 | 12,000 | | 平均每子类样本数 | ~89.73 | | 训练集标签噪声率 | 22.2% – 32.7% | ### 数据集划分 | 划分集 | 样本规模 | 标签质量 | |---|---|---| | 训练集 | 1,036,738 | 网络爬取（含噪声） | | 验证集 | 20,000 | 人工验证（干净标签） | | 测试集 | 20,000 | 人工验证（干净标签） | ### 主类别（共12类） 毛衣、风衣、T恤、衬衫、针织衫、连帽卫衣、夹克、西装、披肩、连衣裙、背心、内衣 ### 子类结构 每个主类别包含**1000个子类**，由三个各含10个选项的属性组合定义： | 属性 | 选项数 | 示例 | |---|---|---| | 颜色 | 10 | 白色、黑色、红色、藏青色、灰色等 | | 材质 | 10 | 棉、羊毛、聚酯纤维、牛仔布等 | | 图案 | 10 | 纯色、条纹、格纹、花卉等 | 搜索查询格式为`"<Color> <Material> <Pattern> <Clothing Type>"`（例如：`"white cotton fisherman sweater"`，即“白色棉质渔夫毛衣”），该查询词同时作为图像搜索的关键词和样本的细粒度标签。 --- ## ADC 构建流程 ADC流程包含三个步骤： **步骤1 — 基于大语言模型的数据集设计** 向GPT-4发送提示词：`"Show me <30–80> ways to describe <Attribute> of <Class>"`，用于枚举每个服装类别的属性选项。对生成的类别进行迭代审核，无需依赖人工领域专业知识。 **步骤2 — 自动标注** 使用组合搜索字符串查询谷歌图像API。每个查询收集约100条顶部搜索结果并自动完成标注。每个查询字符串即为样本的标签，彻底免去手动标注流程。 **步骤3 — 数据整理与清洗** - **算法整理**：使用标签噪声检测方法（例如Simi-Feat / Docta）自动过滤误标注样本，将训练集噪声率从约22.2%降至约10.7%。 - **人机协同（干净划分集）**：对于验证集和测试集，亚马逊机械 Turk（Amazon MTurk）平台上的人工标注员从机器标注的批次中挑选正确样本（每个查询至少20个样本中挑选4个）。仅保留完全符合人机一致标注的样本，纳入干净划分集。 --- ## 基准测试任务 ### 1. 标签噪声检测任务（`Clothing-ADC-Detection`） 该子集包含20000个样本，同时带有噪声标签和干净标签，每张图像由3名亚马逊MTurk工人标注（正确/不确定/错误）。用于基准测试噪声检测算法。 **评估指标：** 检测到的损坏样本的F1分数 | 方法 | F1分数 | |---|---| | CORES | 0.4793 | | 置信度学习（Confident Learning，CL） | 0.4352 | | 深度k近邻（Deep k-NN） | 0.3991 | | Simi-Feat | **0.5721** | ### 2. 噪声标签学习任务（`Clothing-ADC` / `Clothing-ADC-Tiny`） 在完整的带噪声训练集上训练，在干净的留出测试集上评估。同时提供一个小型版本（约5万张训练图像）用于快速实验。 **评估指标：** 干净测试集上的12类别分类准确率 精选实验结果（ResNet-50，20个训练轮次）： | 方法 | 完整数据集 | 小型数据集 | |---|---|---| | 交叉熵（基准方法） | — | — | | 正标签平滑 | ↑ | ↑ | | Taylor CE | **最优** | **最优** | | DivideMix | 具有竞争力 | 具有竞争力 | ### 3. 类别不平衡学习任务（`Clothing-ADC-CLT`） 该数据集的类别长尾分布版本，不平衡比率ρ ∈ {10, 50, 100}。在构建该基准前，使用算法整理方法（Docta + 以学习为目标的整理）移除噪声样本，最终得到约562,263张干净图像。 **评估指标：** δ-最差准确率（在δ=0时为平均准确率，δ→∞时为最差类别准确率的插值指标） | 方法 | ρ=10 (δ=0) | ρ=100 (δ=0) | ρ=10 (δ=∞) | ρ=100 (δ=∞) | |---|---|---|---|---| | 交叉熵 | 57.80 | 30.10 | 0.96 | 0.00 | | 焦点损失（Focal Loss） | 72.70 | 62.28 | 38.12 | 13.44 | | LDAM | 72.50 | 63.25 | 40.90 | 15.69 | | 平衡Softmax | 74.18 | 69.47 | 48.54 | 50.60 | | 对数调整（Logit-Adjust） | **74.08** | **69.44** | 47.45 | 43.26 | | Drops | 73.66 | 67.15 | **50.85** | 32.43 | --- ## 与现有数据集的对比 | 数据集 | 训练集/测试集样本数 | 类别数 | 噪声率（%） | 包含属性 | 自动标注 | 需领域专业知识？ | |---|---|---|---|---|---|---| | iNaturalist | 579k/279k | 54k | ~0 | ✗ | ✗ | ✓ | | WebVision | 2.4M/100k | 1000 | 20 | ✗ | ✓ | ✓ | | ANIMAL-10N | 50k/10k | 10 | 8 | ✗ | ✗ | ✗ | | CIFAR-10N | 50k/10k | 10 | 9–40 | ✗ | ✗ | ✗ | | Food-101N | 75.75k/25.25k | 101 | 18.4 | ✗ | ✗ | ✓ | | Clothing1M | 1M total | 14 | 38.5 | ✗ | ✗ | ✓ | | **Clothing-ADC（本文提出）** | **1M/20k** | **12** | **22.2–32.7** | **12k** | **✓** | **✗** | --- ## 数据字段 每个样本包含以下字段： - `id`：唯一的图像标识符字符串 - `image`：PIL图像（256×256 RGB格式） - `class`：主服装类别字符串（例如：`"Sweater"`，即“毛衣”） - `color`：颜色属性标签（例如：`"white"`，即“白色”） - `material`：材质属性标签（例如：`"cotton"`，即“棉”） - `pattern`：图案属性标签（例如：`"fisherman"`，即“渔夫纹”） --- ## 使用方法 python from datasets import load_dataset # 加载完整数据集 ds = load_dataset("mikelmh025/ClothingADC") # 访问划分集 train = ds["train"] val = ds["validation"] test = ds["test"] # 示例：遍历测试集 for sample in test: image = sample["image"] category = sample["class"] color = sample["color"] material = sample["material"] pattern = sample["pattern"] --- ## 引用 若在研究中使用Clothing-ADC数据集，请引用以下文献： bibtex @article{liu2024adc, title = {Automatic Dataset Construction (ADC): Sample Collection, Data Curation, and Beyond}, author = {Minghao Liu and Zonglin Di and Jiaheng Wei and Zhongruo Wang and Hengxiang Zhang and Ruixuan Xiao and Haoyu Wang and Jinlong Pang and Hao Chen and Ankit Shah and Hongxin Wei and Xinlei He and Zhaowei Zhu and Haobo Wang and Lei Feng and Jindong Wang and James Davis and Yang Liu}, journal = {arXiv preprint arXiv:2408.11338}, year = {2024} } --- ## 许可证 本数据集采用[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)协议发布。图像从谷歌图像搜索收集，仍受其原始来源的版权协议约束。本数据集仅用于研究目的。