Illumination-Robust Object Coordinate Detection by Adopting Pix2Pix GAN for Training Image Generation

Illumination effects often result in error or failure in visual object localization. Whereas ConvNet-based object localization frameworks have shown tremendous robustness to the illumination effect, under some strong illumination effects, the system may still fail. In this paper, the authors propose a data augmentation method utilizing pix2pix GAN for automatic generation of object images under various illumination effects. Upon training for the object localization ConvNet, the generated images are included to enrich the training set for better performance under strong illumination effects. Experimental evidence shows that the accuracy of object coordinate detection can be improved significantly. The proposed framework maintains our concept of “one-shot” where the user only needs to take a basis photo of the target object; the rest of the process including image processing and data annotation are all automated.