NeFr25/TTA_Tidal_Turbine_Assembly_Visual_Dataset

✅ TTA-Sim2Real Dataset

This folder contains a sample version of the TTA-Sim2Real dataset , introduced in our paper "TTA-Sim2Real: A Mixed Real–Synthetic Dataset and Pipeline for Tidal Turbine Assembly Object Detection". The dataset is designed to support object detection in industrial assembly environments, combining real-world footage , controlled captures , and synthetic renderings . This version includes a representative subset for reproducibility and testing purposes. TTA-Sim2Real is a mixed-data object detection dataset designed for sim-to-real research in industrial assembly environments. It includes spontaneous real-world footage , controlled real data captured via cobot-mounted camera , and domain-randomized synthetic images generated using Unity, targeting seven classes related to tidal turbine components at various stages of assembly. The dataset supports reproducibility and benchmarking for vision-based digital twins in manufacturing.

✅ Dataset Card Abstract TTA-Sim2Real is a multi-source object detection dataset specifically designed for sim-to-real transfer in industrial assembly tasks. It contains over 21,000 annotated images across three data types:

-Spontaneous Real Data : Captured from live assembly and disassembly operations, including operator presence with face blurring for privacy. -Controlled Real Data : Structured scenes recorded under uniform lighting and positioning using a cobot-mounted high-resolution camera. -Synthetic Data : 6,000 of auto-labeled images generated using Unity 2022 with domain randomization techniques.

The dataset targets seven object classes representing key turbine components:

-Tidal-turbine -Body-assembled -Body-not-assembled -Hub-assembled -Hub-not-assembled -Rear-cap-assembled -Rear-cap-not-assembled

📁 Folder Structure Overview

dataset/

├ data_access/

│ ├── spontaneous_real_data/ # Unscripted real-world footage ()

│ ├── controlled_real_data/ # Structured scenes from cobot-mounted camera

│ └── synthetic_data/ # Auto-labeled Unity-generated images with domain randomization

├ data_annotation/ # Annotation files and documentation

│ ├── spontaneous_real_data.zip/ # Manual annotations (where available)

│ ├── controlled_real_data.zip/ # Bounding box labels in YOLO format

│ ├── synthetic_data.zip/ # Auto-generated JSON and mask labels

└ README.md # This file

📝 Dataset Description

1. data_access/ Contains sample subsets from three types of data used in our experiments:

🔹 spontaneous_real_data/ Real-world video captured during live assembly operations.

Useful for sim-to-real evaluation and robustness testing.

🔹 controlled_real_data/

Videos captured using a cobot-mounted high-resolution camera.

Contains structured views of turbine components under uniform lighting and angles.

Includes:

Objects of interest only

Objects with small parts

Close-up shots

🔹 synthetic_data/

6,000 auto-labeled images generated using Unity 2022 and Perception Package.

Domain-randomized backgrounds, lighting, and textures.

Includes bounding boxes and segmentation masks in JSON format.

2. data_annotation/

Contains annotation files for training and evaluation:

🔹 spontaneous_real_data.zip/

Semi-automatic annotations where available.

Format:YOLO-compatible .txt files.

🔹 controlled_real_data.zip/

Fully annotated with YOLO-style bounding boxes.

High-quality labels created semi-automatically using CVAT with AI-assisted tools.

🔹 synthetic_data.zip/

Auto-labeled by Unity with accurate bounding boxes and semantic masks.

Includes JSON files with object positions and segmentation labels.