Hardware Tools Complete Dataset
Description
This repository provides a comprehensive, multi-device image dataset covering 24 fine-grained workshop hardware and hand tool categories. Images were captured directly in practical, uncontrolled workshop environments using three distinct smartphone camera modules (OnePlus Nord CE 4, Nothing CMF Phone 2 Pro, and Nothing Phone 2) to capture natural variations in viewpoint, background clutter, ambient/artificial lighting, and optical response. All target objects are annotated with precise 2D bounding boxes in standard COCO JSON format. The master archive (ML_Hardware_Tools_Complete_Dataset.zip) contains: 1. ML_Hardware_Tools_Raw_COCO/: The 4,633 original, unaugmented workshop images partitioned into train, validation, and test splits for custom preprocessing, domain adaptation, and few-shot or self-supervised learning. 2. ML_Hardware_Tools_Augmented_COCO/: The standardized 12,481-image benchmark dataset (10,464 train, 1,004 validation, 1,013 test) containing 14,161 annotated instances used to establish object detection baselines (YOLO variants and RF-DETR). 3. DATASET_DESCRIPTION_VERSION_2.txt: Complete documentation detailing class taxonomy (24 classes), folder organization, annotation schemas, and citation instructions. Source code and benchmark training pipelines are publicly available at: https://github.com/mark-0polo/ML-CODES.git Version 2 Update: The dataset has been refined through duplicate image removal to reduce potential data leakage across dataset splits. Class distribution imbalances have also been addressed to improve dataset balance and consistency.
Files
Steps to reproduce
Download the complete dataset provided in this repository. The dataset has already been created and annotated, so no additional data collection or annotation is required. After downloading, use the provided dataset as input to the model and follow the model implementation to reproduce the reported experiments and results.
Institutions
- United International UniversityDhaka Division, Dhaka