Reproduce “U$^2$Mamba: A Two-level Nested U-structure Mamba for Salient Object Detection” on OpenResearch

About this paper

Junhui Li, Jialu Li, Youshan Zhang

Researchers introduced U2Mamba, a two-level nested U-structure network integrating State Space Models for salient object detection. The model achieved competitive performance on benchmark datasets, securing the lowest MAE of 0.024 on DUTS-TE and HKU-IS, while demonstrating computational efficiency with 41.97M parameters and an inference speed of 24.7 FPS.

Key insights

• The integration of the Mamba state space mechanism within MMUBs effectively models long-range dependencies with linear complexity, offering a more efficient alternative to self-attention for dense prediction tasks like SOD.
• The nested U-structure, combined with MMUBs, successfully balances global contextual understanding from deep layers with the preservation of fine-grained spatial details from shallow layers, crucial for accurate saliency boundaries.
• Hierarchical deep supervision, particularly with the KL divergence loss, is essential for improving inter-level consistency and boundary precision in deep nested architectures, ensuring coherent feature learning across different scales.

Summary by alphaXiv.