PanoDR: Spherical Panorama Diminished Reality for Indoor Scenes

Layout-guided Indoor Panorama Inpainting with Plane-aware Normalization

“…By embedding one style code for individual semantic classes, SEAN shows significant improvement over SPADE and generates the highest quality results. In the context of indoor scenes, Gkitsas et al [9] introduce PanoDR that combines image-toimage translation with a generator to constrain the image inpainting with the underlying scene structure. Percisely, to convert a furnished indoor panorama into its empty counterpart, PanoDR exploits a generator for synthesizing photorealistic image contents where the global layout structure is preserved via an image-to-image translation module.…”

“…Recently, Gkitsas et al [9] introduced PanoDR, a diminished reality-oriented inpainting model for indoor panorama. The main idea is to translate a furnished indoor panorama into its empty counterpart via a network that leverages both a generator and an image-to-image translation module.…”

“…2: Limitations of existing methods. EC [28] and PanoDR [9] fail to align the image structure along the layout boundaries and produce inconsistent blurry image contents in the inpainted regions (red mask).…”

Viewport-Oriented Panoramic Image Inpainting

“…Panoramic images (PIs) have attracted much attention in recent years since they can represent the omnidirectional visual contents. With an increasing need of processing and editing PIs in our daily life, inpainting has become an important task in various PI applications, such as privacy protection in panoramas [1], old panoramic photo restoration, and also PI post-processing for AR [2].…”

“…Recently, several pioneer studies were proposed for panoramic image/video inpainting. Some of them improve the off-the-shelf methods for PIs by applying either semantic conditioning [2,8] or depth information [1]. Besides, some researchers are interested in exploring other projection formats (such as Cubemap in PIINET [9]) to introduce less distortion.…”