DR-CyCADA: Depth Retentive Cycle Consistent Adversarial Domain Adaptation

Abstract

Image depth data can hold a wealth of information for robotic systems that rely on visual input while navigating in unknown environments. Domain adaptation algorithms have enabled robots to transfer knowledge of a known domain to an unfamiliar environment. However, current domain adaptation approaches typically focus on visual appearance, without regard for the structure of an environment. Thus, current approaches can unexpectedly fail when translating between environments. We overcome this mismatch between the original and adapted image by introducing a depth retention loss component to a popular existing domain adaptation model, CyCADA [1]. Our DR-CyCADA approach retains structural components of images and displays promising results in generating high quality images that retain their original structural layout. We evaluate our model via its depth retention loss. We show that DR-CyCADA outperforms similar systems that do not leverage depth retention loss in retaining the structural layout of an environment, demonstrating that our DR-CyCADA model can be used to perform domain adaptation tasks in structurally similar environments and is thus capable of state-of-the-art domain adaptation for robotic systems. 

References: [1] Hoffman, J., Tzeng, E., Park, T., Zhu J., Isola, P., Saenko, K., Efros, A., Darrell, T. (2018, July 10-15). CyCADA: Cycle-Consistent Adversarial Domain Adaptation [Paper Presentation]. International Conference on Machine Learning 35th Annual Meeting, Stockholmsmässan, Sweden. https://proceedings.mlr.press/v80/hoffman18a.html