Xi'an Jiaotong's EQ-VMamba Beats VMamba's Accuracy With 3x Fewer Parameters
A new Mamba variant bakes 90-degree rotation symmetry into the architecture, cutting parameters in half while improving accuracy across vision tasks.
- EQ-VMamba is the first rotation-equivariant visual Mamba, guaranteeing consistent outputs under 90-degree rotations.
- Core innovations: EQ-cross-scan with four symmetric paths and group Mamba blocks sharing weights across orientations.
- Cuts parameters by ~50% versus VMamba while improving ImageNet-100 top-1 accuracy by 0.38-0.78%.
- Beats VMamba by 3-6 mIoU on remote sensing segmentation (LoveDA, ISPRS Potsdam).
- VMamba drops 24 mIoU on rotated ADE20K; EQ-VMamba stays flat under cardinal rotations.
- Code at github.com/zhongchenzhao/EQ-VMamba, supports classification, segmentation, and super-resolution.
Vision Mamba models adapt selective state-space models, or SSMs, to images, processing token sequences with cost that grows linearly with sequence length. Global self-attention typically compares every token pair, so Mamba-style backbones can offer lower scaling costs at high resolutions. Their directional scans also create a weakness: rotating an image by 90 degrees can sharply change the prediction.
Researchers at Xi’an Jiaotong University address that weakness in a new paper introducing EQ-VMamba. The authors describe it as the first visual Mamba architecture with a mathematical guarantee of equivariance under the four cardinal rotations. Their Tiny model reaches 88.58% top-1 accuracy on ImageNet-100 with 10 million parameters, compared with 30 million for VMamba-T.
A fixed scan creates unstable rotations
VMamba converts each two-dimensional feature map into four one-dimensional sequences, processes them along different directional paths, and merges the results. Although the paths cover multiple directions, their token order remains tied to the image’s horizontal and vertical axes. A quarter-turn rotation changes those sequences and the states accumulated along them, while the original cross-scan provides no enforced correspondence between the transformed features.
Data augmentation can expose a model to rotated examples, but it cannot provide an architectural guarantee. The paper reports a 24.17% decline in VMamba-T’s ADE20K segmentation result after rotating the inputs. Such sensitivity is especially relevant for satellite images, microscopy, and other data without a canonical upward direction.
Equivariance keeps orientation intact
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
