Automatic Salient Object Detection for Panoramic Images Using Region Growing and Fixation Prediction Model

Chunbiao Zhu, Kan Huang, Ge LiarXiv 2017 · arXiv

Abstract

Saliency detection—the identification of image regions that naturally attract human attention—is a fundamental task in computer vision. However, traditional models often struggle with panoramic (360-degree) imagery due to extreme aspect ratios and high visual complexity. This paper introduces a novel bottom-up framework that integrates spatial density-based region growing with an eye-fixation prediction model. Evaluated on the newly introduced SalPan dataset, the proposed method demonstrates superior accuracy and computational efficiency compared to twelve state-of-the-art algorithms.

Introduction: The Panoramic Challenge

While saliency detection is well-established for conventional images, panoramic images present unique hurdles. Standard models typically rely on center-surround contrast, a heuristic that fails in high-resolution, wide-field-of-view environments common in VR and AR. Furthermore, the sheer volume of visual information in a 360-degree scene makes it difficult for algorithms to distinguish between a "busy" background and a truly salient object.

Proposed Methodology

The authors propose a dual-pathway framework designed to merge "objectness" (structural proposals) with "fixation" (human visual attention simulation).

The proposed framework consists of dual-stage region growing and fixation prediction pathways, followed by fusion and refinement.

1. Dual-Stage Region Growing

The process begins by segmenting the image into superpixels using the SLIC algorithm. A Density Map is then calculated based on fractal theory (the Holder exponent) to measure intensity variations across scales. This is followed by a "seed growing" process that identifies strong and weak seeds for both foreground and background, iteratively refining object proposals.

2. Fixation Prediction Model

To simulate human visual search, the framework employs an Image Signature model in the frequency domain. This approach is computationally lightweight, making it suitable for the high pixel counts inherent in panoramic imagery.

3. Fusion and Refinement

The outputs from both pathways are fused using Maxima Normalization (MN). Finally, Geodesic Refinement is applied—using the "shortest path" distance on a graph of superpixels—to ensure salient objects are highlighted uniformly while background regions are suppressed.

The SalPan Dataset

To validate their approach, the authors introduced SalPan, a high-quality dataset of 123 panoramic images. Ground truth was established through a user study involving 10 participants and a Gazepoint GP3 Eye Tracker to capture actual human fixation patterns, ensuring the benchmark reflects real-world visual attention.

Experimental Results

The proposed method ("OURS") was benchmarked against 12 state-of-the-art methods. On the SalPan dataset, it achieved the lowest Mean Absolute Error (MAE) of 0.2744 and the highest F-measure (0.7142).

Qualitative comparison of the proposed method against various state-of-the-art saliency detection models across different panoramic scenes.

Key findings from the evaluation include:

  • Density over Contrast: Spatial density patterns proved more reliable than simple color contrast in high-resolution panoramas.
  • Complementary Pathways: Combining "rich" object proposals with "reliable but fuzzy" fixation models successfully offset individual weaknesses.
  • Efficiency: The model processes images in approximately 1.075 seconds, significantly faster than many deep learning alternatives.
  • Small Target Robustness: The model excelled at detecting very small salient objects, a common failure point for conventional models.

Conclusion

The integration of region growing and fixation prediction provides a robust solution for panoramic saliency detection. While the method occasionally struggles with perfectly crisp boundaries in extremely complex scenes, its balance of accuracy and speed makes it highly practical for real-world VR/AR applications. Future work will focus on improving edge-processing techniques and expanding the SalPan dataset to include more diverse scenes.

Reference: Zhu, C., Huang, K., Li, T. H., & Li, G. (2018). Automatic Salient Object Detection for Panoramic Images Using Region Growing and Fixation Prediction Model. arXiv:1710.04071v6 [cs.CV].

Summary updated

BibTeX
@misc{zhu2017automatic,
  author = {Chunbiao Zhu and Kan Huang and Ge Li},
  title = {{Automatic Salient Object Detection for Panoramic Images Using Region Growing and Fixation Prediction Model}},
  year = {2017},
  eprint = {1710.04071},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1710.04071},
}
Figure