Research map / Label-efficient and robust learning / Recognition and long-tail
Enhancing Mamba Decoder with Bidirectional Interaction in Multi-Task Dense Prediction
Mang Cao, Sanping Zhou, Yizhe Li, Ye Deng, Wenli Huang, Le Wang
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add BI-Scan to MTMamba baseline (ablation) | BIM · NYUD-v2 | MTL gain Delta_m: +2.38 -> +4.02 | Improved |
| add MS-Scan to MTMamba baseline (ablation) | BIM · NYUD-v2 | MTL gain Delta_m: +2.38 -> +4.00 | Improved |
| combine BI-Scan and MS-Scan (ablation) | BIM · NYUD-v2 | MTL gain Delta_m: +4.02 -> +4.83 | Improved |
| use both task-first and position-first scan modes (ablation) | BI-Scan · NYUD-v2 | Semseg mIoU: 56.36 / 56.96 -> 57.11 | Improved |
| order TF before PF rather than PF before TF (ablation) | BI-Scan · NYUD-v2 | Semseg mIoU: 56.55 -> 57.11 | Improved |
| use bidirectional instead of unidirectional cross-task scan (ablation) | BI-Scan · NYUD-v2 | overall performance | Improved |
| random task order instead of fixed order (ablation) | BI-Scan · NYUD-v2 | MTL gain Delta_m: +4.02 -> +3.17 | Hurt |
| vary fixed task order (ablation) | BI-Scan · NYUD-v2 | MTL gain Delta_m: +3.72 to +4.20 | No clear effect |
| MS-Scan scales {1,4} versus {1,2} or {1,3} (ablation) | MS-Scan · NYUD-v2 | MTL gain Delta_m: +4.68 / +4.62 -> +4.83 | Improved |
| use more scales across MFRs (Type 3) instead of {1,4} (ablation) | MS-Scan · NYUD-v2 | MTL gain Delta_m: +4.83 -> +4.35 | Hurt |
| dilated sampling in MS-Scan (DBIM) | BIM · NYUD-v2 | FLOPs (G) / Semseg mIoU: 547 -> 495 FLOPs; 57.40 -> 56.41 mIoU | Mixed |
| replace MTMamba decoder interaction with BIM | Swin-Base encoder · NYUD-v2 | Semseg mIoU: 53.62 -> 54.32 | Improved |
Closest papers by meaning
- Demystify Mamba in Vision: A Linear Attention Perspective (NeurIPS 2024 · similarity 0.85)
- MobileMamba: Lightweight Multi-Receptive Visual Mamba Network (CVPR 2025 · similarity 0.84)
- Multi-Task Dense Prediction via Mixture of Low-Rank Experts (CVPR 2024 · similarity 0.84)
- HOIMamba: Efficient Mamba-based Disentangled Progressive Learning for HOI Detection (AAAI 2025 · similarity 0.84)
- Vision Mamba Mender (NeurIPS 2024 · similarity 0.83)
- Achilles' Heel of Mamba: Essential difficulties of the Mamba architecture demonstrated by synthetic data (NeurIPS 2025 · similarity 0.83)
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.
