Introduction
Computer vision for fashion has seen significant progress, yet existing benchmarks like FashionMNIST and DeepFashion remain heavily skewed toward Western attire. Categories such as t-shirts, jeans, and jackets dominate these datasets, leaving a significant gap in the representation of non-Western clothing. Indian ethnic wear—characterized by unique silhouettes, complex draping styles, and intricate patterns—often fails to map correctly to these Western-centric categories.
To address this, the IndoFashion dataset was developed as the first large-scale, publicly available resource specifically for classifying Indian ethnic apparel. With over 106,000 images across 15 distinct categories, it provides the necessary data to train models that can accurately categorize garments like Sarees, Dhotis, and Lehengas.
The IndoFashion Dataset
The core contribution of this work is a robust, fine-grained dataset comprising 106,166 images. Unlike previous ethnic datasets that were either too small or suffered from extreme class imbalance, IndoFashion maintains a substantial average of approximately 7,077 images per class.
The dataset includes 15 popular categories, such as:
- Saree
- Lehenga
- Nehru Jacket
- Sherwani
- Dhoti Pants
- Mojari (traditional footwear)
The data is structured into a training set of 91,166 images, with validation and test sets of 7,500 images each. To ensure a fair and rigorous evaluation, the validation and test sets are perfectly balanced, containing exactly 500 images per category.
Methodology
The researchers employed a systematic approach to data acquisition and model benchmarking:
- Data Collection: Custom web scrapers using BeautifulSoup and Selenium gathered images from major Indian e-commerce platforms (e.g., Flipkart, Myntra, Amazon.in) and Google Images.
- Preprocessing: Images were standardized to a $128 \times 256$ resolution.
- Baseline Models: Standard ImageNet-pretrained architectures—ResNet-18, ResNet-50, and ResNet-101—were evaluated to establish performance baselines.
- Augmentation: To handle the vibrant colors and varied drapes of Indian textiles, the team used a combination of geometric (flipping, rotation) and color (jittering) augmentations.
Key Insights and Results
The benchmarking process revealed several critical findings regarding the classification of ethnic wear:
The Reality of Western Bias
Models trained solely on Western datasets struggle with ethnic outfits because the underlying feature distributions—specifically shapes and textures—are fundamentally different. For instance, the "drape" of a garment is a defining geometric feature in Indian fashion that is largely absent in Western-style clothing.
Performance and Augmentation Synergy
The ResNet-50 model, when combined with both color jittering and horizontal flipping (J+F), emerged as the top performer, achieving a test accuracy of 88.43%. Interestingly, the study found that applying both types of augmentation simultaneously yielded better results than either alone. This suggests that for ethnic wear, the "print" (texture/color) and the "drape" (geometry) are equally vital for identification.
Model Depth vs. Data Quality
While upgrading from ResNet-18 to ResNet-50 provided a clear performance gain, further increasing depth to ResNet-101 showed diminishing returns. This indicates that the quality and diversity of the training data are more significant drivers of accuracy than model complexity for this specific task.
Future Directions
While IndoFashion marks a significant step forward, challenges remain. The confusion matrix indicated semantic overlap between visually similar categories, such as "Women Kurta" versus "Gowns." Future work may involve developing models that can better distinguish between subtle stylistic differences, expanding the scope to include more regional styles, and utilizing the dataset for tasks like landmark detection or personalized fashion recommendation.
By providing a large-scale, balanced dataset, IndoFashion enables the development of more inclusive and effective e-commerce tools for the South Asian market and beyond.
Summary updated