Motif-Sikka-ROI: A Curated and Balanced Image Dataset of Traditional Sikka Ikat Weaving Motifs for Classification and Textile Pattern Analysis

Description

Motif-Sikka-ROI is a curated and balanced image dataset documenting 24 classes of traditional ikat weaving motifs from Sikka Regency, East Nusa Tenggara (NTT), Indonesia. The Sikka ikat weaving tradition is a significant part of the cultural identity of the Sikka people, with each motif carrying distinct symbolic and aesthetic meanings that have been preserved and transmitted across generations. Images were collected through direct field photography at multiple locations, including weavers' homes, fabric shops, traditional markets, and cultural exhibitions in Sikka Regency, using smartphones and digital cameras. This multi-device and multi-location acquisition strategy introduces natural variations in illumination, viewpoint, scale, and background conditions, making the dataset representative of real-world image collection scenarios. Each image was processed using the Roboflow platform, where the primary motif area was manually cropped using Region of Interest (ROI) extraction to remove irrelevant background information and focus on the core weaving pattern. Following ROI extraction, all images underwent a strict manual quality assessment procedure. Images exhibiting blur, excessive occlusion, poor visibility, or insufficient motif detail were excluded from the final dataset. The retained images were standardized to a resolution of 224 × 224 pixels and stored in PNG format to preserve image quality without lossy compression. The dataset contains exactly 180 images for each of the 24 motif classes, resulting in a total of 4,320 ROI images. To improve dataset robustness and support research on illumination-invariant pattern recognition, image enhancement, and domain generalization, the dataset is balanced not only at the class level but also across acquisition conditions. For every motif class, the dataset includes 30 images for each of the following six conditions: morning indoor, morning outdoor, noon indoor, noon outdoor, night indoor, and night outdoor. This condition-balanced design minimizes potential bias arising from unequal lighting and environmental distributions. The balanced composition ensures that no single motif class or acquisition condition dominates during model training, making the dataset suitable for fair benchmarking of machine learning and deep learning algorithms. The dataset is published in an unsplit form to provide maximum flexibility for downstream research tasks, including motif classification, feature extraction, image enhancement, computer vision applications, and the digital preservation of Indonesian cultural heritage.

Steps to reproduce

Step 1 — Field Data Collection • Traditional Sikka ikat sarongs were identified and photographed across multiple locations: weavers' homes, fabric shops, and cultural exhibitions in Sikka Regency, NTT, Indonesia. • Images were captured using a combination of smartphones and digital cameras under various natural and artificial lighting conditions, and at different times of day (morning, afternoon, and evening). • Multiple shots were taken per motif from varying distances and angles to ensure sufficient data variety. Step 2 — Motif Class Definition • A total of 24 distinct traditional motif classes were identified and defined based on visual characteristics and local naming conventions of Sikka ikat weaving. • The traditional local name of each class was used as the class label (see Section 3 for the full list of classes). Step 3—ROI Extraction via Roboflow • All collected images were uploaded to the Roboflow platform (https://roboflow.com). • For each image, a manual Region of Interest (ROI) bounding box was drawn to isolate the primary motif area, removing irrelevant background elements such as hands, surfaces, or surrounding fabric. • Cropped ROI regions were exported as individual image files. Step 4 — Manual Quality Filtering • All ROI-cropped images underwent a manual visual inspection process. • Images that were blurry, poorly lit, partially occluded, or had insufficient pattern visibility were discarded. • Only images with clear, readable motif patterns were retained for the final dataset. Step 5 — Image Standardization • All retained images were resized to a uniform resolution of 224 x 224 pixels. • Images were saved in PNG format (.png) to preserve visual quality without lossy compression. • No additional preprocessing (e.g., color normalization, augmentation, or background removal) was applied, keeping the dataset in its natural state. Step 6 — Class Balancing • Images were organized per class into individual folders. • Each class was balanced to exactly 160 images, ensuring equal representation across all 24 motif classes. • The final dataset contains 4,320 images in total (24 classes x 180 images per class). Step 7 — Dataset Packaging • The dataset was organized into per-class folders following the standard ImageFolder directory structure. • The dataset was published without a predefined train/validation/test split to allow researchers full flexibility in partitioning the data. • This README file was included to document all dataset details and ensure reproducibility.

Institutions

• Universitas Gadjah Mada

  Yogyakarta, Yogyakarta
• Universitas Nusa Nipa Maumere

  East Nusa Tenggara, Maumere

Categories

Licence