Adaptive Multi-Loss GAN for Object Removals in Street-View Scenes

Street-view images often contain personal and sensitive information like pedestrians and vehicles, leading to privacy concerns. It is essential to anonymize these sensitive objects from street-view images without sacrificing image quality. In this study, we introduce a novel synthesis-based object removals in street-view images, named Adaptive Multi-Loss GAN. Our proposed method first extracts area of target objects within a raw street-view image through viewport extraction. Subsequently, a GAN-based network with multiple loss functions is employed to reconstruct the inpaint area. However, a simple weighted average of multi-loss functions may be insufficient and unstable for optimizing the inpaint network. To overcome this issue, we introduce an adaptive weighting strategy. This dynamic approach adjusts the weights of the loss functions based on gradient descent information, facilitating the optimization of the GAN-based network. Extensive experimental evaluations on street-view images datasets demonstrate that our approach achieves the improvement of standard metrics, including Mean-Squared-Error (MSE), Peak Signal-to-Noise Ratio (PSNR)., and Structural Similarity Index (SSIM) over traditional methods that employ a fixed weighted average.