Revisiting Salient Object Detection: Simultaneous Detection, Ranking, and Subitizing of Multiple Salient Objects


Publication

Md Amirul Islam*, Mahmoud Kalash*, Neil D. B. Bruce
Revisiting Salient Object Detection: Simultaneous Detection, Ranking, and Subitizing of Multiple Salient Objects, CVPR, 2018. (Oral)
[BibTeX] [PDF]
@inproceedings{islamsal18,
  author = {A. Islam, M. Kalash, N. D. B. Bruce},
  title = {Revisiting Salient Object Detection: Simultaneous Detection, Ranking, and Subitizing of Multiple Salient Objects},
  booktitle = {Computer Vision and Pattern Recognition (CVPR)},
  year = {2018}
}

Results


Salient Object Detection:

The table shows the quantitative comparison of methods including AUC, max F-measure (higher is better), median F-measure, average F-measure, MAE (lower is better), and SOR (higher is better).



Predicted salient object regions for the Pascal-S dataset. Each row shows outputs corresponding to different algorithms designed for the salient object detection/segmentation task.




Saliency Ranking:

Below are the qualitative comparison of the state-of-the-art approaches designed for salient object detection. Relative rank is indicated by the assigned color.




Examining the Nested Relative Salience Stack:

Quantitative comparison (AUC & Fm) with state-of-the-art methods across all ground truth thresholds, each corresponding to agreement among a specific number participants. Best and second best scores are shown in red and blue respectively.



Below figure shows the top three principal components as an RGB image where the first principal component (which captures the most variance across layers) is mapped to the R-channel, the second component is mapped to the G-channel and so forth.