SalNet360: Saliency maps for omni-directional images with CNN

作者：

Highlights：

• We use deep learning to predict saliency in ODI images.

• We extend CNN networks for saliency prediction on traditional 2D images to ODIs.

• Distortions in an ODI are reduced by projecting it to square patches.

• We use spherical coordinates allowing the network to infer any directional biases.

摘要

•We use deep learning to predict saliency in ODI images.•We extend CNN networks for saliency prediction on traditional 2D images to ODIs.•Distortions in an ODI are reduced by projecting it to square patches.•We use spherical coordinates allowing the network to infer any directional biases.

论文关键词：Saliency,Omnidirectional Image (ODI),Convolutional Neural Network (CNN),Virtual Reality (VR)

论文评审过程：Received 14 September 2017, Revised 28 March 2018, Accepted 7 May 2018, Available online 12 May 2018, Version of Record 30 October 2018.

论文官网地址：https://doi.org/10.1016/j.image.2018.05.005