Eunsu Baek

Computer vision applications predict on digital images acquired by a camera from physical scenes through light. However, conventional robustness benchmarks rely on perturbations in digitized images, diverging from distribution shifts occurring in the image acquisition process. To bridge this gap, we introduce a new distribution shift dataset, ImageNet-ES, comprising variations in environmental and camera sensor factors by directly capturing 202k images with a real camera in a controllable testbed. With the new dataset, we evaluate out-of-distribution (OOD) detection and model robustness. We find that existing OOD detection methods do not cope with the covariate shifts in ImageNet-ES, implying that the definition and detection of OOD should be revisited to embrace real-world distribution shifts. We also observe that the model becomes more robust in both ImageNet-C and -ES by learning environment and sensor variations in addition to existing digital augmentations. Lastly, our results suggest that effective shift mitigation via camera sensor control can significantly improve performance without increasing model size. With these findings, our benchmark may aid future research on robustness, OOD, and camera sensor control for computer vision. Our code and dataset are available at this https://github.com/Edw2n/ImageNet-ES

We present MIRROR, an on-device video virtual try-on (VTO) system that provides realistic, private, and rapid experiences in mobile clothes shopping. Despite recent advancements in generative adversarial networks (GANs) for VTO, designing MIRROR involves two challenges: (1) data discrepancy due to restricted training data that miss various poses, body sizes, and backgrounds and (2) local computation overhead that uses up 24% of battery for converting only a single video. To alleviate the problems, we propose a generalizable VTO GAN that not only discerns intricate human body semantics but also captures domain-invariant features without requiring additional training data. In addition, we craft lightweight, reliable clothes/pose-tracking that generates refined pixel-wise warping flow without neural-net computation. As a holistic system, MIRROR integrates the new VTO GAN and tracking method with meticulous pre/post-processing, operating in two distinct phases (on/offline). Our results on Android smartphones and real-world user videos show that compared to a cutting-edge VTO GAN, MIRROR achieves 6.5x better accuracy with 20.1x faster video conversion and 16.9x less energy consumption.