PISA: pixelwise image saliency by aggregating complementary appearance contrast measures with edge-preserving...

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Accurate pixel-level saliency detection is crucial for computer vision tasks like image segmentation and object recognition.
• Existing methods often rely on color-only or superpixel-based approaches, limiting their detail preservation and spatial coherence.

Purpose of the Study:

• To introduce a unified framework, Pixelwise Image Saliency Aggregating (PISA), for generating pixel-accurate and fine-grained saliency maps.
• To overcome the limitations of previous saliency detection methods by incorporating diverse bottom-up cues and spatial priors.

Main Methods:

• PISA aggregates multiple saliency cues (e.g., color contrast, structure contrast) within a global context, incorporating spatial priors.
• A neighborhood consistency constraint and energy minimization formulation are used, solved via cost-volume filtering for smooth, edge-aware results.
• A faster version of PISA utilizes gradient-driven image subsampling for improved runtime efficiency.

Main Results:

• PISA generates spatially coherent, detail-preserving, and pixel-accurate saliency maps.
• The cost-volume filtering approach ensures smooth saliency level assignment while preserving structural details.
• Experiments demonstrate that PISA significantly outperforms existing state-of-the-art saliency detection methods.
• A new dataset of 800 commodity images was created for saliency detection evaluation.

Conclusions:

• PISA provides a robust and effective framework for pixel-accurate saliency detection, outperforming previous approaches.
• The method's ability to integrate multiple cues and its efficient implementation make it suitable for various vision applications.
• The introduction of a new dataset facilitates further research and benchmarking in saliency detection.