3Harvard Medical School, 4Xi'an Jiaotong University
Selective state space models (SSMs), such as Mamba, highly excel at capturing long-range dependencies in 1D sequential data, while their applications to 2D vision tasks still face challenges. Current visual SSMs often convert images into 1D sequences and employ various scanning patterns to incorporate local spatial dependencies. However, these methods are limited in effectively capturing the complex image spatial structures and the increased computational cost caused by the lengthened scanning paths. To address these limitations, we propose Spatial-Mamba, a novel approach that establishes neighborhood connectivity directly in the state space. Instead of relying solely on sequential state transitions, we introduce a structure-aware state fusion equation, which leverages dilated convolutions to capture image spatial structural dependencies, significantly enhancing the flow of visual contextual information. Spatial-Mamba proceeds in three stages: initial state computation in a unidirectional scan, spatial context acquisition through structure-aware state fusion, and final state computation using the observation equation. Our theoretical analysis shows that Spatial-Mamba unifies the original Mamba and linear attention under the same matrix multiplication framework, providing a deeper understanding of our method. Experimental results demonstrate that Spatial-Mamba, even with a single scan, attains or surpasses the state-of-the-art SSM-based models in image classification, detection and segmentation.
-
Clone repo
git clone https://github.com/EdwardChasel/Spatial-Mamba.git cd Spatial-Mamba -
Create and activate a new conda environment
conda create -n Spatial-Mamba python=3.10 conda activate Spatial-Mamba
-
Install dependent packages
pip install --upgrade pip pip install -r requirements.txt cd kernels/selective_scan && pip install . cd kernels/dwconv2d && python3 setup.py install --user
-
Dependencies for detection and segmentation (optional)
pip install mmengine==0.10.1 mmcv==2.1.0 opencv-python-headless ftfy regex pip install mmdet==3.3.0 mmsegmentation==1.2.2 mmpretrain==1.2.0
ImageNet-1k Image Classification
COCO Object Detection and Instance Segmentation
| backbone | method | schedule | mAP (box/mask) | #param | FLOPs | download |
|---|---|---|---|---|---|---|
| Spatial-Mamba-T | Mask R-CNN | 1x | 47.6 / 42.9 | 46M | 261G | ckpt | config |
| Spatial-Mamba-S | Mask R-CNN | 1x | 49.2 / 44.0 | 63M | 315G | ckpt | config |
| Spatial-Mamba-B | Mask R-CNN | 1x | 50.4 / 45.1 | 115M | 494G | ckpt | config |
| Spatial-Mamba-T | Mask R-CNN | 3x | 49.3 / 43.8 | 46M | 261G | ckpt | config |
| Spatial-Mamba-S | Mask R-CNN | 3x | 50.5 / 44.6 | 63M | 315G | ckpt | config |
ADE20K Semantic Segmentation
-
ImageNet is an image database organized according to the WordNet hierarchy. Download and extract ImageNet train and val images from http://image-net.org/. Organize the data into the following directory structure:
imagenet/ βββ train/ β βββ n01440764/ (Example synset ID) β β βββ image1.JPEG β β βββ image2.JPEG β β βββ ... β βββ n01443537/ (Another synset ID) β β βββ ... β βββ ... βββ val/ βββ n01440764/ (Example synset ID) β βββ image1.JPEG β βββ ... βββ ... -
COCO is a large-scale object detection, segmentation, and captioning dataset. Please visit http://cocodataset.org/ for more information, including for the data, paper, and tutorials. COCO API also provides a concise and efficient way to process the data.
-
ADE20K is composed of more than 27K images from the SUN and Places databases. Please visit https://ade20k.csail.mit.edu/ for more information and see the GitHub Repository for an overview of how to access and explore ADE20K.
-
Image Classification
To train Spatial-Mamba models for classification on ImageNet, use the following commands for different configurations:
cd classification python -m torch.distributed.launch --nnodes=1 --node_rank=0 --nproc_per_node=8 --master_addr="127.0.0.1" --master_port=29501 main.py --cfg </path/to/config> --batch-size 128 --data-path </path/of/dataset> --output /tmp
To evaluate the performance with pre-trained weights:
cd classification python -m torch.distributed.launch --nnodes=1 --node_rank=0 --nproc_per_node=1 --master_addr="127.0.0.1" --master_port=29501 main.py --cfg </path/to/config> --batch-size 128 --data-path </path/of/dataset> --output /tmp --pretrained </path/of/checkpoint>
-
Detection and Segmentation
To train with mmdetection or mmsegmentation:
cd detection bash ./tools/dist_train.sh </path/to/detection/config> 8
cd segmentation bash ./tools/dist_train.sh </path/to/segmentation/config> 8
To evaluate with mmdetection or mmsegmentation:
cd detection bash ./tools/dist_test.sh </path/to/detection/config> </path/to/detection/checkpoint> 1
cd segmentation bash ./tools/dist_test.sh </path/to/segmentation/config> </path/to/segmentation/checkpoint> 1
use
--ttato get themIoU(ms)in segmentation.
@inproceedings{
xiao2025spatialmamba,
title={Spatial-Mamba: Effective Visual State Space Models via Structure-Aware State Fusion},
author={Chaodong Xiao and Minghan Li and Zhengqiang Zhang and Deyu Meng and Lei Zhang},
booktitle={The Thirteenth International Conference on Learning Representations},
year={2025}
}This project is largely based on Mamba, VMamba, MLLA, Swin-Transformer, RepLKNet and OpenMMLab. We are truly grateful for their excellent work.
This project is released under the Apache 2.0 license.