Research map / Detection and segmentation / CNN features
Mamba YOLO: A Simple Baseline for Object Detection with State Space Model
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 Mamba YOLO, a real-time object detector whose ODMamba backbone uses SS2D state-space blocks plus a multi-branch RG Block, trained from scratch without ImageNet pretraining, and evaluated on COCO.
Problem
Transformer-based detectors achieve strong accuracy but have quadratic self-attention cost, and existing SSM-based vision backbones typically rely on large-scale pretraining; the paper targets a real-time, pretraining-free SSM detector.
Approach
Introduces ODSSBlock (SS2D token mixing plus a gated RG Block for channel mixing) as a YOLO backbone and neck replacement for C2f, with a Vision Clue Merge downsampling that keeps more information for SSM scanning and a Simple Stem for patch embedding.
Key results
- Mamba YOLO-T reaches 44.5% AP on COCO val with 5.8M params, 13.2G FLOPs and 1.5 ms latency on a 4090 GPU in FP16.
- Mamba YOLO-B achieves 49.1% AP at 19.1M params and 2.2 ms, versus 49.0% for PPYOLOE-M (which uses pretraining) at 4.3 ms.
- Ablation on Mamba YOLO-T shows ODSSBlock raises AP from 37.3 to 43.1 and the full ODSSBlock + RG Block + Clue Merge reaches 44.5.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ODSSBlock (ablation) | Mamba YOLO-T · MSCOCO val | AP: 37. 3 -> 43. 1 | Helped |
| Add RG Block (ablation) | Mamba YOLO-T · MSCOCO val | AP: 37. 3 -> 37. 9 | No clear effect |
| Add Clue Merge (ablation) | Mamba YOLO-T · MSCOCO val | AP: 37. 3 -> 35. 6 | Hurt |
| Add ODSSBlock and RG Block (ablation) | Mamba YOLO-T · MSCOCO val | AP: 37. 3 -> 44. 1 | Helped |
| Add Clue Merge to ODSSBlock and RG Block (ablation) | Mamba YOLO-T · MSCOCO val | AP: 44. 1 -> 44. 5 | Helped |
| Add DW-Conv to original MLP (ablation) | Mamba YOLO-T RG module · MSCOCO val | AP: 43. 0 -> 43. 2 | No clear effect |
| Add residual-connected DW-Conv to original MLP (ablation) | Mamba YOLO-T RG module · MSCOCO val | AP: 43. 0 -> 43. 3 | No clear effect |
| Use gated MLP (ablation) | Mamba YOLO-T RG module · MSCOCO val | AP: 43. 0 -> 44. 0 | Helped |
| Use RG Block instead of original MLP (ablation) | Mamba YOLO-T · MSCOCO val | AP: 43. 0 -> 44. 5 | Helped |
| Change backbone ODSSBlock repetitions to [9, 3, 3, 3] (ablation) | Mamba YOLO-T · MSCOCO val | AP: 44. 5 -> 44. 0 | Hurt |
| Change backbone ODSSBlock repetitions to [3, 9, 3, 3] (ablation) | Mamba YOLO-T · MSCOCO val | AP: 44. 5 -> 43. 4 | Hurt |
| Change backbone ODSSBlock repetitions to [3, 3, 9, 3] (ablation) | Mamba YOLO-T · MSCOCO val | AP: 44. 5 -> 43. 6 | Hurt |
Surprising findings
- Mamba YOLO-T trained from scratch outperforms several competitors that use ImageNet or large-scale detection pretraining.
Limitations
- Comparisons rely on reported numbers and some baselines use extra self-distillation or pretraining, complicating fair comparison.
- Ablation of design choices is performed only on the Tiny model.
- Speed claims depend on a specific GPU and TensorRT setup.
Experimental setup
Datasets: MSCOCO
Benchmarks: MSCOCO val
Metrics: AP AP50 AP75 APS APM APL FLOPs latency params
Baselines: YOLOv6-3.0 YOLOv8 YOLO-MS DAMO-YOLO PPYOLOE Gold-YOLO DINO-R50
Builds on: Mamba VMamba YOLOv8 Gated MLP PAFPN
Closest papers by meaning
- YOLOv10: Real-Time End-to-End Object Detection (NeurIPS 2024 · similarity 0.90)
- Samba: A Unified Mamba-based Framework for General Salient Object Detection (CVPR 2025 · similarity 0.89)
- Vision Mamba Mender (NeurIPS 2024 · similarity 0.89)
- Demystify Mamba in Vision: A Linear Attention Perspective (NeurIPS 2024 · similarity 0.89)
- MambaOut: Do We Really Need Mamba for Vision? (CVPR 2025 · similarity 0.89)
- SMamba: Sparse Mamba for Event-based Object Detection (AAAI 2025 · similarity 0.88)
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.