Saliency Detection: A Boolean Map Approach

Jianming Zhang, Stan SclaroffICCV 2013 · OpenReview

Abstract

A novel Boolean Map based Saliency (BMS) model is proposed. An image is characterized by a set of binary images, which are generated by randomly thresholding the image's color channels. Based on a Gestalt principle of figure-ground segregation, BMS computes saliency maps by analyzing the topological structure of Boolean maps. BMS is simple to implement and efficient to run. Despite its simplicity, BMS consistently achieves state-of-the-art performance compared with ten leading methods on five eye tracking datasets. Furthermore, BMS is also shown to be advantageous in salient object detection.

Introduction

Saliency detection aims to predict human visual attention by identifying the most important regions in a scene. Historically, models have relied on local contrast or statistical rarity to highlight salient features. However, these methods often fail in complex environments where high-contrast background elements (like tree branches) can distract from the primary subject. The Boolean Map based Saliency (BMS) model addresses this by leveraging the Gestalt principle of surroundedness to perform effective figure-ground segregation.

Methodology

The BMS model utilizes a streamlined pipeline that prioritizes topological structure over complex statistical modeling:

  1. Boolean Map Generation: The input image (in CIE Lab color space) is decomposed into a set of binary "Boolean maps" by thresholding color channels at random levels.
  2. Attention Map Computation: For each map, the algorithm identifies "surrounded" regions using a Flood Fill approach. Any region not reachable from the image borders is considered a "figure" and is activated.
  3. Aggregation and Post-processing: These individual maps are averaged to create a mean attention map, which is then dilated and blurred (Gaussian) to produce the final saliency output.

Unlike contemporary models that require multi-scale pyramids or heavy machine learning, BMS is scale-invariant and computationally efficient due to its reliance on basic binary topology.

Key Results

BMS was evaluated against ten leading saliency models across five benchmark eye-tracking datasets, including MIT and Toronto. It consistently achieved the highest average Shuffled-AUC, demonstrating superior robustness across diverse visual scenes.

Comparison of saliency maps: (a) Input image and human fixations; (b) Saliency maps from AIM, LG, and the proposed BMS model.

As illustrated in the figure above, BMS effectively isolates the salient object (the bird) while suppressing background noise from the trees—a task where competing models like AIM and LG often struggle. Furthermore, on the ASD salient object detection dataset, BMS produced clean object boundaries, proving its utility for computer vision tasks like image segmentation.

Discussion and Conclusion

The success of BMS suggests that figure-ground segregation, often considered a high-level cognitive process, can be effectively modeled as a low-level, bottom-up visual cue. Its simplicity allows for high-speed processing, handling a 600x400 image in approximately 0.38 seconds.

While highly effective, the model currently faces limitations when salient objects touch the image border, as the Flood Fill may "leak" into the object. Future research could involve incorporating additional features like texture, motion, and depth, or exploring other Gestalt principles like symmetry to further refine the saliency maps.

Summary updated

Abstract

A novel Boolean Map based Saliency (BMS) model is proposed. An image is characterized by a set of binary images, which are generated by randomly thresholding the image's color channels. Based on a Gestalt principle of figure-ground segregation, BMS computes saliency maps by analyzing the topological structure of Boolean maps. BMS is simple to implement and efficient to run. Despite its simplicity, BMS consistently achieves state-of-the-art performance compared with ten leading methods on five eye tracking datasets. Furthermore, BMS is also shown to be advantageous in salient object detection.
BibTeX
@inproceedings{zhang2013saliency,
  author = {Jianming Zhang and Stan Sclaroff},
  title = {{Saliency Detection: A Boolean Map Approach}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2013},
  url = {https://openreview.net/forum?id=rJEDm-M_ZH},
}
Figure