Research map / Detection and segmentation / CNN features
GroupMamba: Efficient Group-Based Visual 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 GroupMamba, a hierarchical vision SSM built from a Modulated Group Mamba layer (four channel groups with four-direction VSSS scans plus Channel Affinity Modulation) trained with a distillation loss from RegNetY-16G, reaching 83.3% ImageNet top-1 with 23M params.
Problem
Visual state-space models (Mamba-style) are parameter- and compute-heavy as channel count grows and can be unstable when scaled to large sizes for image classification and dense prediction.
Approach
Split channels into four groups, each scanned by a VSSS block in a different spatial direction (LR, RL, TB, BT), recombine them with a Channel Affinity Modulation SE-style recalibration to restore cross-group interaction, and stabilize large models with a teacher-student distillation loss.
Key results
- GroupMamba-T achieves 83.3% ImageNet-1K top-1 with 23M parameters and 4.5G FLOPs, outperforming VMamba-T (82.5%) with 26% fewer parameters.
- GroupMamba-B reaches 84.5% top-1 with 57M parameters, exceeding VMamba-B by 0.6% with 36% fewer parameters.
- GroupMamba-T gets 47.6 box AP and 42.9 mask AP on COCO with Mask R-CNN 1x, and 48.6 mIoU (SS) on ADE20K.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GroupMamba-T instead of VMamba-T | Image classification · ImageNet-1K | top-1 accuracy: 82.5 -> 83.3 | Helped |
| use GroupMamba-B instead of VMamba-B | Image classification · ImageNet-1K | top-1 accuracy: 83.9 -> 84.5 | Helped |
| use GroupMamba-T instead of VMamba-T | Mask R-CNN · MS-COCO 2017 | box AP: 47.4 -> 47.6 | Helped |
| use GroupMamba-T instead of VMamba-T | Mask R-CNN · MS-COCO 2017 | mask AP: 42.7 -> 42.9 | Helped |
| use GroupMamba-T instead of VMamba-T | UperNet · ADE20K single-scale | mIoU: 48.3 -> 48.6 | Helped |
| use GroupMamba-T instead of VMamba-T | UperNet · ADE20K multi-scale | mIoU: 48.6 -> 49.2 | Helped |
| replace 4-D scanning with unidirectional 1-D scanning across N/4 channels in four directions (ablation) | GroupMamba-T · ImageNet-1K | throughput: 803 -> 1125 | Helped |
| replace 4-D scanning with unidirectional 1-D scanning across N/4 channels in four directions (ablation) | GroupMamba-T · ImageNet-1K | top-1 accuracy: 82.30% -> 82.20% | Hurt |
| add Channel Affinity Modulation (ablation) | GroupMamba-T with 1-D scanning · ImageNet-1K | top-1 accuracy: 82.20% -> 82.50% | Helped |
| add Channel Affinity Modulation (ablation) | GroupMamba-T with 1-D scanning · ImageNet-1K | throughput: 1125 -> 1069 | Hurt |
| add distillation-based loss (ablation) | GroupMamba-T with 1-D scanning and CAM · ImageNet-1K | top-1 accuracy: 82.50% -> 83.30% | Helped |
| add distillation-based loss (ablation) | GroupMamba-T with 1-D scanning and CAM · ImageNet-1K | throughput: 1069 -> 1069 | No clear effect |
Limitations
- Evaluation limited to image classification, detection, and segmentation; video and time-series untested.
- Distillation stability benefit is reported mainly in supplementary material.
Experimental setup
Datasets: ImageNet-1K MS-COCO 2017 ADE20K
Benchmarks: ImageNet-1K MS-COCO 2017 ADE20K
Metrics: top-1 accuracy box AP mask AP mIoU parameters FLOPs throughput
Baselines: VMamba LocalVMamba Vim EfficientVMamba Swin Transformer ConvNeXt DeiT RegNetY EfficientNet ResNet-50
Builds on: Mamba VMamba Vision Mamba Swin Transformer Squeeze-and-Excitation
Closest papers by meaning
- Multi-Scale VMamba: Hierarchy in Hierarchy Visual State Space Model (NeurIPS 2024 · similarity 0.91)
- EfficientVMamba: Atrous Selective Scan for Light Weight Visual Mamba (AAAI 2025 · similarity 0.90)
- Spatial-Mamba: Effective Visual State Space Models via Structure-Aware State Fusion (ICLR 2025 · similarity 0.90)
- QuadMamba: Learning Quadtree-based Selective Scan for Visual State Space Model (NeurIPS 2024 · similarity 0.90)
- Vision Mamba Mender (NeurIPS 2024 · similarity 0.88)
- VMamba: Visual State Space Model (NeurIPS 2024 · 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.