Towards Fully-Synthetic Training for Industrial Applications

Abstract

This paper proposes a scalable approach for synthetic image generation of industrial objects leveraging Blender for image rendering. In addition to common components in synthetic image generation research, three novel features are presented: First, we model relations between target objects and randomly apply those during scene generation (Object Relation Modelling (ORM)). Second, we extend the idea of distractors and create Object-alike Distractors (OAD), resembling the textural appearance (i.e. material and size) of target objects. And third, we propose a Mixed-lighting Illumination (MLI), combining global and local light sources to automatically create a diverse illumination of the scene. In addition to the image generation approach we create an industry-centered dataset for evaluation purposes. Experiments show, that our approach enables fully synthetic training of object detectors for industrial use-cases. Moreover, an ablation study provides evidence on the performance boost in object detection when using our novel features.

ROSUS 2023 - Računalniška obdelava slik in njena uporaba v Sloveniji 2023: Zbornik 17. strokovne konference

“…generative adversarial networks) (Dewi et al, 2022) in difuzijski modeli (Pinaya et al, 2022;Akrout et al, 2023) v nasprotju z obdobjem pred tem, ko se je večino sintetičnih slik ustvarilo s programi za grafično upodabljanje. Večina metod (Li et al, 2019) uporablja sintetične slike kot podaljšek učne množice, ki je sestavljena samo iz resničnih slik, nekaj pa jih zgradi celotno učno množico iz sintetičnih slik (Mayershofer et al, 2021;Mayer et al, 2016) in nato naredi premik domene z uporabo resničnih slik za doučenje (Wang et al, 2019;Richter et al, 2016) nevronske mreže.…”

“…V industrijski domeni (Mayershofer et al, 2021;Eversberg et al, 2021;Abou et al, 2022;Quattrocchi et al, 2022) pogosto pride do uporabe sintetičnih slik, saj je pogosto zajemanje in označevanje velike količine slik drago in zamudno. Velika večina pristopov ustvari sintetične slike na podoben način.…”

Self-Supervised Skin Lesion Segmentation: An Annotation-Free Approach

“…Supervised deep learning methods applied to images have demonstrated their potential in extracting image features to support various image analysis tasks. Many of the techniques that have shown success in logistics applications, such as image segmentation for parcel [5,6] and packaging recognition [7,8], have the potential to be applied in the medical image domain as well. These techniques leverage underlying algorithms and rely on abundant labeled data which is costly [9], underscoring their efficacy and promise [10,11].…”

Generating Synthetic Training Data for Assembly Processes

“…The second main approach generating synthetic training data is Domain Randomization (DR) [14,21,29]. The idea is to make reality appear as just another synthetic modification of the training images.…”