PSC522 Dataset
Public PublishedThis dataset is part of the publication "A PLM-Integrated automated synthetic image generation pipeline for object detection" presented at the PLM conference 2025 in Sevilla. The paper presents an automated synthetic image data generation pipeline aimed at streamlining the training process of object detection models supporting manual assembly processes. By automating the rendering of images from CAD models instead of relying on manually created physical product images, the pipeline enables dataset creation in earlier phases of the product lifecycle while also significantly reducing manual effort. This approach enhances the accessibility for finetuned object detection model development. The pipeline integrates two core components: object similarity analysis and synthetic image generation. The similarity analysis groups visually similar objects into unified classes for the object detection model, reducing confusion during detection. The image generation process can be augmented with contextual information from virtual 3D workplace scenes, thereby significantly mitigating the sim-to-real gap. The pipeline is accessed via a REST API, enabling seamless integration with PLM systems for automated retrieval of CAD models and workplace scene data. A workflow manager orchestrates interactions between the user, the PLM system, and the generation pipeline. The effectiveness of the system is validated by evaluating object detection models trained on the synthetically generated datasets against real-world images, demonstrating its potential to improve detection accuracy and robustness in industrial environments. By publishing the data used, we aim to strengthen the traceability of the results obtained and that we can encourage further research in this field.
- Title
- PSC522 Dataset
- Creator or Contributor
- Resource type
- Keyword
- License
- Description
- This dataset is part of the publication "A PLM-Integrated automated synthetic image generation pipeline for object detection" presented at the PLM conference 2025 in Sevilla. The paper presents an automated synthetic image data generation pipeline aimed at streamlining the training process of object detection models supporting manual assembly processes. By automating the rendering of images from CAD models instead of relying on manually created physical product images, the pipeline enables dataset creation in earlier phases of the product lifecycle while also significantly reducing manual effort. This approach enhances the accessibility for finetuned object detection model development. The pipeline integrates two core components: object similarity analysis and synthetic image generation. The similarity analysis groups visually similar objects into unified classes for the object detection model, reducing confusion during detection. The image generation process can be augmented with contextual information from virtual 3D workplace scenes, thereby significantly mitigating the sim-to-real gap. The pipeline is accessed via a REST API, enabling seamless integration with PLM systems for automated retrieval of CAD models and workplace scene data. A workflow manager orchestrates interactions between the user, the PLM system, and the generation pipeline. The effectiveness of the system is validated by evaluating object detection models trained on the synthetically generated datasets against real-world images, demonstrating its potential to improve detection accuracy and robustness in industrial environments. By publishing the data used, we aim to strengthen the traceability of the results obtained and that we can encourage further research in this field.
- Date
- Language
- Identifier
- Publisher
- Last modified
- 2025-06-05
Relations
Items
| Thumbnail | Title | Date Uploaded | Visibility | Actions |
|---|---|---|---|---|
|
PSC522_Train_Dataset.zip | 2025-06-03 | Public | Download |
|
|
PSC522_Test_Dataset.zip | 2025-06-05 | Public | Download |