Simple Baselines Can Fool 360° Saliency Metrics
Abstract
Evaluating a model’s capacity to predict human fixations in 360° scenes is a challenging task. 360° saliency requires different assumptions compared to 2D as a result of the way the saliency maps are collected and pre-processed to account for the difference in statistical bias (Equator vs Center bias). However, the same classical metrics from the 2D saliency literature are typically used to evaluate 360° models. In this paper, we show that a simple constant predictor, i.e. the average map across Salient360 and Sitzman training sets can fool existing metrics and achieve results on par with specialized models. Thus, we propose a new probabilistic metric based on the independent Bernoullis assumption that is more suited to the 360° saliency task.
Cite
Text
Djilali et al. "Simple Baselines Can Fool 360° Saliency Metrics." IEEE/CVF International Conference on Computer Vision Workshops, 2021. doi:10.1109/ICCVW54120.2021.00418Markdown
[Djilali et al. "Simple Baselines Can Fool 360° Saliency Metrics." IEEE/CVF International Conference on Computer Vision Workshops, 2021.](https://mlanthology.org/iccvw/2021/djilali2021iccvw-simple/) doi:10.1109/ICCVW54120.2021.00418BibTeX
@inproceedings{djilali2021iccvw-simple,
title = {{Simple Baselines Can Fool 360° Saliency Metrics}},
author = {Djilali, Yasser Abdelaziz Dahou and McGuinness, Kevin and O'Connor, Noel E.},
booktitle = {IEEE/CVF International Conference on Computer Vision Workshops},
year = {2021},
pages = {3743-3749},
doi = {10.1109/ICCVW54120.2021.00418},
url = {https://mlanthology.org/iccvw/2021/djilali2021iccvw-simple/}
}