Research map / Detection and segmentation / CNN features
Samba: A Unified Mamba-based Framework for General Salient Object Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Samba, a unified pure-Mamba framework for RGB, RGB-D, RGB-T, video and RGB-D video salient object detection, using a spatial neighboring scanning strategy and context-aware upsampling.
Problem
Salient object detection across RGB, RGB-D, RGB-T, video and RGB-D video tasks is dominated by CNNs (limited receptive fields) and transformers (quadratic cost); existing Mamba scanning strategies break spatial continuity of salient patches and Mamba-based decoders use non-learnable or context-blind
Approach
A saliency-guided Mamba block whose spatial neighboring scanning (SNS) orders salient patches row by row along an approximate shortest path, plus a context-aware upsampling that pairs shallow and deep patches as subsequences fed to S6 blocks, and a multi-modal fusion Mamba for combining modalities.
Key results
- Samba reports DUTS Sm 0.932 / Fm 0.930 for RGB SOD, outperforming VSCode-S (0.926 / 0.922) with 49.59M params.
- On RGB-D VSOD ViDSOD, Samba reaches Sm 0.923 and Em 0.944 versus ATFNet's 0.875 and 0.911.
- Across 21 datasets in five SOD tasks, Samba achieves best results with lower MACs than most transformer-based competitors, though VSCode-S received more training data.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add SGMB (ablation) | Samba · DUTS | Fm: 0.916 -> 0.930 | Helped |
| use SG-SS2D instead of SS2D (ablation) | Samba · DUTS | Fm: 0.919 -> 0.930 | Helped |
| use SNS instead of Z-pattern scanning (ablation) | Samba · DUTS | Fm: 0.924 -> 0.930 | Helped |
| use SNS instead of diagonal scanning (ablation) | Samba · DUTS | Fm: 0.927 -> 0.930 | Helped |
| use SNS instead of S-pattern scanning (ablation) | Samba · DUTS | Fm: 0.928 -> 0.930 | Helped |
| use three scanning paths with varying directions instead of copies of the initial path (ablation) | Samba · DUTS | Fm: 0.926 -> 0.930 | Helped |
| use CAU instead of nearest-neighbor interpolation (ablation) | Samba · DUTS | Fm: 0.923 -> 0.930 | Helped |
| use CAU instead of DUpsampling (ablation) | Samba · DUTS | Fm: 0.918 -> 0.930 | Helped |
| use CAU instead of DySample (ablation) | Samba · DUTS | Fm: 0.924 -> 0.930 | Helped |
| use the original CAU patch-pairing order instead of a one-patch shift (ablation) | Samba · DUTS | Fm: 0.928 -> 0.930 | Helped |
| use the original CAU patch-pairing order instead of a three-patch shift (ablation) | Samba · DUTS | Fm: 0.926 -> 0.930 | Helped |
| use the original CAU patch-pairing order instead of a five-patch shift (ablation) | Samba · DUTS | Fm: 0.922 -> 0.930 | Helped |
Surprising findings
- Compared to VSCode-S, which was trained jointly on all SOD subtasks, Samba is trained per sub-task separately yet still performs best on the reported datasets.
Limitations
- SNS adds extra computation to determine per-row scanning direction compared with fixed scan patterns.
- Samba is trained separately per sub-task, whereas VSCode-S uses joint training data from all subtasks, so comparisons are not data-matched.
- RGB-D VSOD has no benchmark training set, so models were trained and tested separately on each dataset.
Experimental setup
Datasets: DUTS NJUD NLPR DUTLF-D VT5000 DAVIS DAVSOD RDVS DVisal ViDSOD
Benchmarks: DUTS DUT-O HKU-IS PASCAL-S ECSSD NJUD NLPR SIP STERE DUTLF-D VT821 VT1000
Metrics: structure-measure (Sm) maximum F-measure (Fm) enhanced-alignment measure (Em) MACs Params
Baselines: VSCode-S VST++ ICON-S EBM MENet SwinNet-B CATNet CPNet ATFNet DCTNet+
Builds on: VMamba Mamba DySample DUpsampling VSCode
Closest papers by meaning
- Mamba YOLO: A Simple Baseline for Object Detection with State Space Model (AAAI 2025 · similarity 0.89)
- Salient Object Detection With Pyramid Attention and Salient Edges (CVPR 2019 · similarity 0.89)
- Recurrently Aggregating Deep Features for Salient Object Detection (AAAI 2018 · similarity 0.88)
- DVSOD: RGB-D Video Salient Object Detection (NeurIPS 2023 · similarity 0.88)
- Spatial-Mamba: Effective Visual State Space Models via Structure-Aware State Fusion (ICLR 2025 · similarity 0.87)
- Demystify Mamba in Vision: A Linear Attention Perspective (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.