# MRGAN360: Multi-stage Recurrent Generative Adversarial Network for 360 Degree Image Saliency Prediction

**Authors:** Pan Gao, Xinlang Chen, Rong Quan, Wei Xiang  
**arXiv:** [2303.08525](https://arxiv.org/abs/2303.08525)  
**Date:** 2023-03-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.08525v1/

## Key points

- 360 degree image content is popular in consumer and industrial applications for its immersive and interactive experience
- Predicting saliency for 360 degree images is challenging due to high resolutions and spherical viewing ranges
- Researchers propose a multi-stage recurrent generative adversarial network called MRGAN360 to address this issue
- MRGAN360 leverages human visual perception processes to predict saliency maps stage by stage, with a recurrent neural network modeling correlations and a discriminator supervising output at each stage
- Weights are shared among all stages for a lightweight architecture that is computationally cheap
- Experiments show that MRGAN360 outperforms state-of-the-art models in prediction accuracy and model size.

## Summary

360 degree image content has become increasingly popular in both consumer and industrial applications due to its ability to provide an immersive and interactive experience. Compared to planar 2D images, predicting saliency for 360 degree images is more challenging due to their high resolutions and spherical viewing ranges. To address this issue, a team of researchers comprising Pan Gao, Xinlang Chen, Rong Quan, and Wei Xiang propose a novel multi-stage recurrent generative adversarial network called MRGAN360 in their paper titled "MRGAN360: Multi-stage Recurrent Generative Adversarial Network for 360 Degree Image Saliency Prediction". The proposed model leverages insights from human visual perception processes and predicts saliency maps stage by stage. At each stage, the model takes as input the original image and the output of the previous stage and outputs a more accurate saliency map. A recurrent neural network is employed among adjacent prediction stages to model their correlations while a discriminator at the end of each stage supervises the output saliency map. Additionally, weights are shared among all stages to obtain a lightweight architecture that is computationally cheap. Experiments conducted by the authors demonstrate that MRGAN360 outperforms state-of-the-art models in terms of both prediction accuracy and model size. In summary, MRGAN360 is a promising approach for predicting saliency maps for omnidirectional images due to its ability to achieve high accuracy while being computationally efficient.