# Cross-View Panorama Image Synthesis

**Authors:** Songsong Wu, Hao Tang, Xiao-Yuan Jing, Haifeng Zhao, Jianjun Qian, Nicu Sebe, Yan Yan  
**arXiv:** [2203.11832](https://arxiv.org/abs/2203.11832)  
**Date:** 2022-03-22  
**Summary page:** https://www.summarizepaper.com/en/paper/2203.11832v1/

## Key points

- Paper title: "Cross-View Panorama Image Synthesis"
- Authors: Songsong Wu, Hao Tang, Xiao-Yuan Jing, Haifeng Zhao, Jianjun Qian, Nicu Sebe, Yan Yan
- Objective: Generate ground-view panorama images from top-view aerial images
- Proposed solution: PanoGAN framework
- Adversarial feedback GAN framework
- Consists of adversarial feedback module and dual branch discrimination strategy
- Generator produces target panorama image and segmentation map from aerial image input
- Feedback loop refines intermediate representations for improved generation performance
- Pixel-segmentation alignment mechanism ensures high-fidelity and semantic consistency in generated images
- Experimental results show PanoGAN outperforms existing approaches in producing high-quality detailed panorama images

## Summary

The paper titled "Cross-View Panorama Image Synthesis" by Songsong Wu, Hao Tang, Xiao-Yuan Jing, Haifeng Zhao, Jianjun Qian, Nicu Sebe, and Yan Yan addresses the challenging task of generating ground-view panorama images from top-view aerial images. This is a difficult problem due to the significant differences in viewpoint between the two image domains. To tackle this issue, the authors propose a novel adversarial feedback GAN framework called PanoGAN. This framework consists of an adversarial feedback module and a dual branch discrimination strategy. The aerial image is first input into the generator to produce a target panorama image and its corresponding segmentation map. This approach helps train the model with layout semantics. Then, feature responses from the discriminator are fed back to the generator through the adversarial feedback module to refine intermediate representations in an iterative process that improves generation performance over time. To ensure high-fidelity and semantic consistency in the generated panorama image, the authors introduce a pixel-segmentation alignment mechanism within the dual branch discrimination strategy. This mechanism promotes collaboration between the generator and discriminator and leads to improved overall output quality. Experimental results on challenging cross-view image datasets demonstrate that PanoGAN outperforms existing approaches by producing high-quality panorama images with more detailed features. The paper has been accepted for publication in IEEE Transactions on Multimedia.