Gyeong-Moon Park — Page 3

arXiv2022

IntereStyle: Encoding an Interest Region for Robust StyleGAN Inversion

Recently, manipulation of real-world images has been highly elaborated along with the development of Generative Adversarial Networks (GANs) and corresponding encoders, which embed real-world images into the latent space. However, designing encoders of GAN still remains a challenging task due to the trade-off between distortion and perception. In this paper, we point out that the existing encoders try to lower the distortion not only on the interest region, e.g., human facial region but also on the uninterest region, e.g., background patterns and obstacles. However, most uninterest regions in real-world images are located at out-of-distribution (OOD), which are infeasible to be ideally reconstructed by generative models. Moreover, we empirically find that the uninterest region overlapped with the interest region can mangle the original feature of the interest region, e.g., a microphone overlapped with a facial region is inverted into the white beard. As a result, lowering the distortion of the whole image while maintaining the perceptual quality is very challenging. To overcome this trade-off, we propose a simple yet effective encoder training scheme, coined IntereStyle, which facilitates encoding by focusing on the interest region. IntereStyle steers the encoder to disentangle the encodings of the interest and uninterest regions. To this end, we filter the information of the uninterest region iteratively to regulate the negative impact of the uninterest region. We demonstrate that IntereStyle achieves both lower distortion and higher perceptual quality compared to the existing state-of-the-art encoders. Especially, our model robustly conserves features of the original images, which shows the robust image editing and style mixing results. We will release our code with the pre-trained model after the review.

cs.CV↗

arXiv2021

Continual Unsupervised Domain Adaptation for Semantic Segmentation

Unsupervised Domain Adaptation (UDA) for semantic segmentation has been favorably applied to real-world scenarios in which pixel-level labels are hard to be obtained. In most of the existing UDA methods, all target data are assumed to be introduced simultaneously. Yet, the data are usually presented sequentially in the real world. Moreover, Continual UDA, which deals with more practical scenarios with multiple target domains in the continual learning setting, has not been actively explored. In this light, we propose Continual UDA for semantic segmentation based on a newly designed Expanding Target-specific Memory (ETM) framework. Our novel ETM framework contains Target-specific Memory (TM) for each target domain to alleviate catastrophic forgetting. Furthermore, a proposed Double Hinge Adversarial (DHA) loss leads the network to produce better UDA performance overall. Our design of the TM and training objectives let the semantic segmentation network adapt to the current target domain while preserving the knowledge learned on previous target domains. The model with the proposed framework outperforms other state-of-the-art models in continual learning settings on standard benchmarks such as GTA5, SYNTHIA, CityScapes, IDD, and Cross-City datasets. The source code is available at https://github.com/joonh-kim/ETM.

cs.CV↗