360-aware saliency estimation with conventional image saliency predictors
This work explores saliency prediction for panoramic 360°-scenes stored as equirectangular images, using exclusively regular “flat” image saliency predictors. The simple equirectangular projection causes severe distortions in the resulting image, which need to be compensated for sensible saliency pr...
Saved in:
Published in | Signal processing. Image communication Vol. 69; pp. 43 - 52 |
---|---|
Main Authors | , |
Format | Journal Article |
Language | English |
Published |
Amsterdam
Elsevier B.V
01.11.2018
Elsevier BV |
Subjects | |
Online Access | Get full text |
Cover
Loading…
Summary: | This work explores saliency prediction for panoramic 360°-scenes stored as equirectangular images, using exclusively regular “flat” image saliency predictors. The simple equirectangular projection causes severe distortions in the resulting image, which need to be compensated for sensible saliency prediction in all viewports. To address this and other arising issues, we propose several ways of interpreting equirectangular images and analyse how these affect the quality of the resulting saliency maps. We perform our experiments with three popular conventional saliency predictors and achieve excellent results on the “Salient360!” Grand Challenge data set (ranked 1st among the blind-test submissions in the Head–Eye Saliency Prediction track).
[Display omitted]
•Applied to equirectangular content directly, 2D saliency predictors perform poorly.•Different input image transformations can eliminate various artefacts.•Combining them with 2D image saliency models won the Salient360! Grand Challenge. |
---|---|
ISSN: | 0923-5965 1879-2677 |
DOI: | 10.1016/j.image.2018.03.013 |