Deep learning approaches for improving robustness in real-time 3D-object positioning and manipulation in severe lighting conditions.

In this article, we address the critical role of illumination in visual positioning and propose an illumination augmentation scheme to improve positioning accuracy under various illumination conditions. We use pix2pix GAN to train various illumination models that generate illumination images for 3D objects. These images are then used to augment the training of deep learning-based visual positioning models, such as OneShot. Our experimental results demonstrate a significant improvement in the positioning accuracy of the object manipulation experiments conducted by an automated visual-servo manipulator. Illumination augmentations increased the success rate from 41 to 89% under severe illuminations and from 80 to 97% under ordinary illuminations. The proposed illumination augmentation scheme offers a practical solution for improving the visual positioning of objects in diverse and changing manufacturing environments. [ABSTRACT FROM AUTHOR]