tomrb/yolov8to_data

Dataset Card for YOLOv8-TO_Data

This dataset contains the training and testing sets for the YOLOv8-TO paper.

Dataset Description

• Created by: Thomas Rochefort-Beaudoin
• License: MIT
• Datasets
  • MMC Dataset Description: The MMC (Minimum Compliance) dataset is derived using the MMC method as the basis for the training dataset, where the segmentation labels are generated from black-and-white density projections obtained via a Heaviside projection. Split: 80% training, 10% validation, 10% testing Usage: Model training and evaluation
  • Random Assembly Dataset Description: This dataset consists of assemblies composed of randomly distributed components, generated to allow for cost-effective training data production. The design variables sampled randomly define the segmentation labels for detection and regression tasks. Usage: Training only
  • SIMP Dataset Description: Generated using the Solid Isotropic Material with Penalization (SIMP) method, this dataset includes 2000 TO structures, allowing to test the model's capability as a general post-processing tool. Samples: 2000 Usage: Testing
  • Low Volume Fraction SIMP Dataset (SIMP5%) Description: Comprising 2000 random SIMP samples with a low volume fraction (5%), this dataset features thin structures that simulate "truss-like" properties suitable for comparison against skeletonization approaches. Samples: 2000 Usage: Testing
  • Out-of-Distribution (OOD) Dataset Description: This dataset includes 4 TO structure images from the literature, featuring complex structures like 2D femur structures and cantilever beams optimized under various constraints to test the model's generalization capabilities. Samples: 4 Usage: Testing

Dataset Sources

• Repository: https://github.com/COSIM-Lab/YOLOv8-TO
• Paper: https://arxiv.org/pdf/2404.18763
• Demo: https://huggingface.co/spaces/tomrb/YOLOv8-TO

Dataset Structure

IMPORTANT: The dataset currently has the design variables of each component in position 1 to 7 in each row. These design variables are currently ignored by the YOLOv8-TO library and are artifacts of when we were trying to do regression directly on the design variables.

Citation

BibTeX:

@misc{rochefortbeaudoin2024density,
      title={From Density to Geometry: YOLOv8 Instance Segmentation for Reverse Engineering of Optimized Structures}, 
      author={Thomas Rochefort-Beaudoin and Aurelian Vadean and Sofiane Achiche and Niels Aage},
      year={2024},
      eprint={2404.18763},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}