Toward Stable Co-Saliency Detection and Object Co-Segmentation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Accurate co-saliency detection (CoSOD) and co-segmentation (CoSEG) require effective modeling of inter-image relationships.
• Recurrent Neural Networks (RNNs) have been used but suffer from order-sensitivity, impacting model stability.
• Existing methods struggle with stable and robustly capturing relationships across image groups.

Purpose of the Study:

• To propose a novel model for simultaneous stable CoSOD and CoSEG.
• To address the order-sensitivity limitations inherent in RNN-based approaches.
• To enhance the stability and accuracy of inter-image relation modeling in CoSOD and CoSEG tasks.

Main Methods:

• Introduction of a Multi-Path Stable Recurrent Unit (MSRU) incorporating Dummy Orders Mechanisms (DOM) and a Recurrent Unit (RU).
• Development of a Cross-Order Contrastive Loss (COCL) to mitigate order-sensitivity by aligning feature embeddings from different input orders.
• Validation using established CoSOD datasets (CoCA, CoSOD3k, Cosal2015, iCoseg, MSRC) and CoSEG datasets (Internet, iCoseg, PASCAL-VOC).

Main Results:

• The proposed MSRU effectively captures robust inter-image relations while reducing order-sensitivity.
• The COCL further enhances stability by ensuring consistent feature representations regardless of input order.
• Experimental results demonstrate superior performance compared to state-of-the-art (SOTA) methods on multiple benchmark datasets.

Conclusions:

• The novel model achieves stable and accurate simultaneous CoSOD and CoSEG.
• The MSRU and COCL components effectively address the critical challenge of order-sensitivity in inter-image relation modeling.
• The proposed approach represents a significant advancement over existing SOTA methods in the field.