Research map / Architectures and efficiency / Attention and transformers
JamMa: Ultra-lightweight Local Feature Matching with Joint Mamba
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 JamMa, an ultra-lightweight Mamba-based semi-dense feature matcher whose JEGO joint scan-merge strategy gives mutual, efficient, global, omnidirectional interaction between two images at linear cost.
Problem
Transformer-based semi-dense feature matchers model long-range dependencies with quadratic cost, causing high training cost and inference latency; visual Mamba scan strategies are designed for single images and lack cross-image interaction.
Approach
A joint scan that alternates between the two images' coarse features (for high-frequency mutual interaction), combined with skip scanning in four directions (keeping total sequence length N), and a 3x3 gated convolutional aggregator on a balanced receptive field, yielding global omnidirectional features with linear complexity.
Key results
- On MegaDepth relative pose, JamMa reaches AUC@5/10/20 of 64.1/77.4/86.5 with 5.7M params, 202.9 GFLOPs-scale efficiency metrics and 59.9 ms runtime, achieving the best average efficiency-performance rank of 3 among 14 methods.
- On HPatches homography estimation, JamMa gets 68.1 AUC@3px, best among compared methods, while using only 36% of ASpanFormer's parameters.
- Replacing joint scan with sequential scan drops pose AUC by about 2.3-2.6 points; JEGO beats EVMamba scan by about 2.2-2.6 points and is faster than VMamba scan (68.3 ms vs 61.8 ms).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace joint scan with sequential scan (ablation) | JamMa · MegaDepth (15-epoch ablation) | AUC@5°: 64.5 -> 62.2 | Hurt |
| add joint scan instead of sequential scan (ablation) | JamMa · MegaDepth (15-epoch ablation) | AUC@10°: 74.7 -> 77.3 | Helped |
| add the aggregator (ablation) | JamMa · MegaDepth (15-epoch ablation) | AUC@5°: 62.3 -> 64.5 | Helped |
| add the aggregator (ablation) | JamMa · MegaDepth (15-epoch ablation) | AUC@20°: 84.3 -> 86.3 | Helped |
| replace JEGO scan with EVMamba scan (ablation) | JamMa · MegaDepth (15-epoch ablation) | AUC@5°: 64.5 -> 61.9 | Hurt |
| replace JEGO scan with VMamba scan (ablation) | JamMa · MegaDepth (15-epoch ablation) | AUC@5°: 64.5 -> 64.1 | Hurt |
| replace Mamba with vanilla attention (ablation) | JamMa · MegaDepth (15-epoch ablation) | memory feasibility | Hurt |
| replace Mamba with linear attention (ablation) | JamMa · MegaDepth (15-epoch ablation) | AUC@5°: +0.3% | Helped |
| add Mamba instead of linear attention (ablation) | JamMa · MegaDepth (15-epoch ablation) | Time (ms): 24.3 -> 3.2 | Helped |
| add cross-image interaction (ablation) | JamMa · MegaDepth (15-epoch ablation) | AUC@5°: 60.1 -> 64.5 | Helped |
| replace ConvNeXt with XFeat encoder (ablation) | JamMa · MegaDepth | AUC@5°: 64.5 -> 61.1 | Hurt |
| replace ConvNeXt with SuperPoint encoder (ablation) | JamMa · MegaDepth | AUC@5°: 64.5 -> 61.8 | Hurt |
Surprising findings
- Zero-shot transfer to indoor ScanNet does not fully transfer; JamMa lags LoFTR there despite competitive outdoor results, which the authors attribute to its small parameter count.
- Coarse layers are no longer the bottleneck after Joint Mamba; coarse matching softmax on the similarity matrix becomes the new runtime bottleneck.
Limitations
- Weaker generalization to indoor scenes (ScanNet) due to small parameter count.
- Coarse matching softmax becomes the runtime bottleneck after Joint Mamba.
Experimental setup
Datasets: MegaDepth
Benchmarks: MegaDepth HPatches ScanNet
Metrics: AUC of pose error @5/10/20 degrees AUC of homography corner error @3/5/10 px parameters FLOPs runtime (ms) average rank
Baselines: XFeat SuperPoint+SuperGlue SuperPoint+LightGlue DKM RoMa LoFTR ASpanFormer ELoFTR MatchFormer RCM
Builds on: Mamba VMamba EVMamba XoFTR LoFTR ConvNeXt V2
Closest papers by meaning
- ETO:Efficient Transformer-based Local Feature Matching by Organizing Multiple Homography Hypotheses (NeurIPS 2024 · similarity 0.87)
- EfficientVMamba: Atrous Selective Scan for Light Weight Visual Mamba (AAAI 2025 · similarity 0.87)
- Efficient LoFTR: Semi-Dense Local Feature Matching with Sparse-Like Speed (CVPR 2024 · similarity 0.85)
- MobileMamba: Lightweight Multi-Receptive Visual Mamba Network (CVPR 2025 · similarity 0.85)
- XFeat: Accelerated Features for Lightweight Image Matching (CVPR 2024 · similarity 0.85)
- HomoMatcher: Achieving Dense Feature Matching with Semi-Dense Efficiency by Homography Estimation (AAAI 2025 · similarity 0.85)
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.