[Paper Review] The iMaterialist Fashion Attribute Dataset

Sheng Guo, Weilin Huang+7 more|arXiv (Cornell University)|Jun 13, 2019

Generative Adversarial Networks and Image Synthesis42 references4 citations

TL;DR

This paper introduces the iMaterialist Fashion Attribute (iFashion-Attribute) dataset, the first million-scale, multi-label, fine-grained fashion image dataset with expert-annotated attributes across 228 classes in 8 groups. Trained on this dataset, deep CNNs achieve state-of-the-art transfer learning performance on fashion benchmarks like DeepFashion, outperforming models pre-trained on ImageNet or Clothes-1M.

ABSTRACT

Large-scale image databases such as ImageNet have significantly advanced image classification and other visual recognition tasks. However much of these datasets are constructed only for single-label and coarse object-level classification. For real-world applications, multiple labels and fine-grained categories are often needed, yet very few such datasets exist publicly, especially those of large-scale and high quality. In this work, we contribute to the community a new dataset called iMaterialist Fashion Attribute (iFashion-Attribute) to address this problem in the fashion domain. The dataset is constructed from over one million fashion images with a label space that includes 8 groups of 228 fine-grained attributes in total. Each image is annotated by experts with multiple, high-quality fashion attributes. The result is the first known million-scale multi-label and fine-grained image dataset. We conduct extensive experiments and provide baseline results with modern deep Convolutional Neural Networks (CNNs). Additionally, we demonstrate models pre-trained on iFashion-Attribute achieve superior transfer learning performance on fashion related tasks compared with pre-training from ImageNet or other fashion datasets. Data is available at: https://github.com/visipedia/imat_fashion_comp

Motivation & Objective

• Address the lack of large-scale, high-quality, multi-label, and fine-grained image datasets in the fashion domain.
• Create a benchmark dataset with over one million fashion images annotated with 228 fine-grained attributes across 8 categories.
• Enable research in multi-label image recognition and fine-grained visual classification using real-world fashion data.
• Demonstrate the superiority of pre-training on iFashion-Attribute for transfer learning on downstream fashion recognition tasks.
• Provide strong baseline results using modern CNNs to establish a new performance standard for fashion attribute prediction.

Proposed method

• Curated a dataset of over one million fashion images from diverse sources, with expert annotations for 228 fine-grained attributes across 8 groups: category, color, gender, material, neckline, pattern, sleeve, and style.
• Employed a multi-stage annotation pipeline combining automated filtering, expert labeling, and quality control to ensure high label accuracy and consistency.
• Used state-of-the-art deep Convolutional Neural Networks (CNNs), including Inception-BN, for training and evaluation on multi-label attribute prediction.
• Evaluated transfer learning performance by fine-tuning models pre-trained on iFashion-Attribute on the DeepFashion and Clothes-50K benchmarks.
• Applied standard training protocols with cross-entropy loss and standard data augmentation, without specialized noise-robust training techniques.
• Reported performance using precision, recall, and top-k accuracy metrics on validation sets, comparing with ImageNet and Clothes-1M pre-training.

Experimental results

Research questions

• RQ1 Can a large-scale, multi-label, fine-grained fashion attribute dataset improve performance in multi-label image recognition compared to existing benchmarks?
• RQ2 How does pre-training on iFashion-Attribute compare to pre-training on ImageNet or Clothes-1M in terms of transfer learning performance on fashion-specific tasks?
• RQ3 To what extent does the fine-grained, high-intra-class diversity of fashion attributes challenge standard CNNs and existing recognition benchmarks?
• RQ4 Does the inclusion of expert-annotated, multi-label attributes in a million-scale dataset lead to better generalization and transferability in downstream tasks?
• RQ5 Can a simple, standard CNN training protocol on iFashion-Attribute outperform complex, noise-robust methods trained on noisier datasets like Clothes-1M?

Key findings

• The iFashion-Attribute dataset contains 913,857 training images with 228 fine-grained attributes across 8 groups, with an average of 5.0 labels per image.
• Models pre-trained on iFashion-Attribute achieved 88.2% Top-1 and 93.3% Top-5 accuracy on the DeepFashion validation set, surpassing models pre-trained on ImageNet (87.3% and 92.9%) and Clothes-1M (85.9% and 91.9%).
• On the Clothes-50K benchmark, the iFashion-Attribute pre-trained model achieved 80.5% accuracy, comparable to specialized noise-robust methods trained on the noisier Clothes-1M dataset.
• The dataset's high-quality, expert-annotated labels and fine-grained nature significantly improve model generalization and transfer performance, even without complex data-cleaning techniques.
• The results demonstrate that iFashion-Attribute is a superior pre-training source for fashion-related vision tasks, outperforming both ImageNet and Clothes-1M in transfer learning benchmarks.
• Ablation experiments confirmed that label quality and scale are critical: a simple model trained on iFashion-Attribute outperformed more complex models trained on noisier data, highlighting the value of clean, expert-annotated data.