Learning Discriminative Representation for Co-Salient Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Co-salient object detection (CoSOD) identifies common salient objects across multiple images.
• Existing CoSOD methods often separate feature extraction and interimage relation modeling, limiting performance in complex scenes.

Purpose of the Study:

• To propose a novel CoSOD framework that unifies feature extraction and interimage relation modeling.
• To improve the discriminative power of features for co-salient objects.

Main Methods:

• Introduced an Early Token Interaction Module (ETIM) for simultaneous feature extraction and interimage information interaction.
• Developed a Pixel-to-Group Contrastive (PGC) learning method to enhance feature discriminability without extra modules.
• Proposed a streamlined network architecture comprising a backbone with ETIM and a decoder.

Main Results:

• The proposed framework achieved state-of-the-art performance on CoCA, CoSOD3k, and Cosal2015 benchmarks.
• The unified approach and PGC learning effectively improved the detection of co-salient objects, especially in cluttered environments.
• The method demonstrated superior performance compared to current leading CoSOD models.

Conclusions:

• The novel CoSOD framework effectively unifies feature extraction and relation modeling, leading to enhanced performance.
• The ETIM and PGC learning contribute significantly to improving feature discriminability and overall CoSOD accuracy.
• The proposed method represents a significant advancement in co-salient object detection.