Salient Object Detection via Dynamic Scale Routing.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Salient Object Detection (SOD) models leverage deep learning for multi-scale feature representation.
• Current SOD models often use fixed kernel sizes, which may not be optimal for varying object scales.

Purpose of the Study:

• To introduce a novel dynamic scale routing mechanism for salient object detection.
• To improve SOD performance by adapting feature extraction to object sizes.

Main Methods:

• Proposed Dynamic Pyramid Convolution (DPConv) to dynamically select kernel sizes based on input.
• Developed a self-adaptive bidirectional decoder to complement the DPConv encoder.
• Implemented a generic plug-in module compatible with existing feature backbones.

Main Results:

• The proposed DPNet enhances salient object detection performance.
• Dynamic scale routing enables scale-aware inference by routing between feature scales.
• Achieved state-of-the-art (SOTA) results on SOD benchmarks.

Conclusions:

• Dynamic scale routing is a promising approach for improving salient object detection.
• The DPNet architecture offers a flexible and effective solution for scale-aware SOD.
• Publicly available code and dataset facilitate further research and application.