Stage-wise Salient Object Detection in 360° Omnidirectional Image via Object-level Semantical Saliency Ranking

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Salient Object Detection (SOD) in 2D images is well-researched, but 360° omnidirectional image SOD faces significant challenges.
• Existing 360° SOD methods are hindered by insufficient training data, visual distortions causing feature gaps, and complex multi-tasking approaches.
• The infeasibility of traditional stage-wise training due to feature discrepancies further exacerbates data scarcity issues in 360° SOD.

Purpose of the Study:

• To address the limitations in 360° omnidirectional image salient object detection.
• To propose a novel multi-stage task decomposition for 360° SOD.
• To introduce a method for ranking object-level semantical saliency to improve viewpoint and object localization.

Main Methods:

• Decomposed the complex 360° SOD problem into a sequence of simpler sub-problems, each requiring smaller training datasets.
• Developed a technique to rank "object-level semantical saliency" for precise localization.
• Introduced the 360-SSOD dataset, featuring 1,105 annotated 360° images with balanced semantic distribution.

Main Results:

• The proposed multi-stage approach effectively mitigates training data shortages and visual distortion issues.
• The object-level semantical saliency ranking enables accurate identification of salient viewpoints and objects.
• Experimental comparisons against 13 state-of-the-art methods demonstrate the superiority of the proposed method.

Conclusions:

• The multi-stage task decomposition is a viable strategy for tackling complex problems like 360° SOD with limited data.
• The novel dataset and ranking mechanism significantly advance the field of 360° salient object detection.
• The research provides a robust and efficient solution for salient object detection in omnidirectional imagery.