Synthetic Object Recognition Dataset for Industries

Abstract

Smart robots in factories highly depend on Computer Vision (CV) tasks, e.g. object detection and recognition, to perceive their surroundings and react accordingly. These CV tasks can be performed after training deep learning (DL) models on large annotated datasets. In an industrial setting, acquiring and annotating such datasets is challenging because it is time-consuming, prone to human error, and limited by several privacy and security regulations. In this study, we propose a synthetic industrial dataset for object detection purposes created using NVIDIA Omniverse. The dataset consists of S industrial assets in 32 scenarios and 200,000 photo-realistic rendered images that are annotated with accurate bounding boxes. For evaluation purposes, multiple object detectors were trained with synthetic data to infer on real images captured inside a factory. Accuracy values higher than 50% and up to 100% were reported for most of the considered assets.