2S-ODIS: Two-Stage Omni-Directional Image Synthesis by Geometric Distortion Correction

TL;DR

The paper tackles the challenge of generating high-quality omni-directional ERP images from NFoV inputs with reduced training cost. It introduces 2S-ODIS, a two-stage method that uses a pre-trained VQGAN without fine-tuning to produce a global coarse ERP image in stage one and then refines it by synthesizing 26 NFoV views in stage two to correct geometric distortions, all under a non-autoregressive MaskGIT framework. This approach significantly lowers training time (from about 14 days to around 4 days) and speeds up inference, while achieving superior quantitative metrics (FID, IS, LPIPS) compared with baselines like OmniDreamer; it also demonstrates robust performance across inpainting, outpainting, and multi-view conditioning scenarios. The method leverages MaxViT-based architectures, a cosine-mask scheduling strategy, and distance-based NFoV fusion to deliver globally coherent and locally detailed omni-directional imagery, with practical implications for VR, SNS, and mapping applications and potential extensions to other omni-directional tasks.

Abstract

Omni-directional images have been increasingly used in various applications, including virtual reality and SNS (Social Networking Services). However, their availability is comparatively limited in contrast to normal field of view (NFoV) images, since specialized cameras are required to take omni-directional images. Consequently, several methods have been proposed based on generative adversarial networks (GAN) to synthesize omni-directional images, but these approaches have shown difficulties in training of the models, due to instability and/or significant time consumption in the training. To address these problems, this paper proposes a novel omni-directional image synthesis method, 2S-ODIS (Two-Stage Omni-Directional Image Synthesis), which generated high-quality omni-directional images but drastically reduced the training time. This was realized by utilizing the VQGAN (Vector Quantized GAN) model pre-trained on a large-scale NFoV image database such as ImageNet without fine-tuning. Since this pre-trained model does not represent distortions of omni-directional images in the equi-rectangular projection (ERP), it cannot be applied directly to the omni-directional image synthesis in ERP. Therefore, two-stage structure was adopted to first create a global coarse image in ERP and then refine the image by integrating multiple local NFoV images in the higher resolution to compensate the distortions in ERP, both of which are based on the pre-trained VQGAN model. As a result, the proposed method, 2S-ODIS, achieved the reduction of the training time from 14 days in OmniDreamer to four days in higher image quality.

Figures (13)

• Figure 1: Overview of advantages of proposed method, 2S-ODIS. OmniDreamer omnidreamer requires 14 days for training of the model, including 1-week training of the VQGAN model. In contrast, the proposed method only required 4 days for the training of the model since no training of VQGAN model was required.
• Figure 2: Qualitative comparison of omni-directional image reconstruction using pre-trained VQGAN encoder and decoder. Omni-directional Image: original omni-directional image, Reconstructed in ERP: reconstructed in equirectangular projection, Reconstructed in Extracted Images: reconstructed by integrating multiple NFoV images in different directions. By extracting NFoV images, an omni-directional image can be correctly reconstructed without distortions.
• Figure 3: Diagram of the proposed method. (a)Inference, (b)Training.
• Figure 4: Structure of the proposed method. The range of attention differs between the high-resolution and low-resolution models.
• Figure 5: Examples of conditional image in training. These images are generated from omni-directional images in ERP by randomly masking.