Research map / Video / Video segmentation and tracking
MambaSCI: Efficient Mamba-UNet for Quad-Bayer Patterned Video Snapshot Compressive Imaging
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 MambaSCI, a Mamba-UNet with spatial-temporal Mamba, edge-detail and channel-attention residual blocks for reconstructing quad-Bayer patterned color video snapshot compressive imaging.
Problem
Existing color video SCI reconstruction algorithms are designed for traditional Bayer patterns and produce color distortion and poor demosaicing when applied to quad-Bayer measurements, while transformer-based E2E methods are computationally heavy.
Approach
Combines a non-symmetric UNet with Residual-Mamba-Blocks (STMamba for parallel spatial forward/backward and temporal scanning, an EDR module of linear layers plus DWConv, and a CA module) and learnable residual scales, with a 3D-conv color reconstruction head that avoids explicit demosaicing.
Key results
- MambaSCI-B reaches 35.70 dB average PSNR on 6 mid-scale simulated color videos, versus 32.48 dB for STFormer-B and 32.35 dB for EfficientSCI-B, using 6.11M params and 556.89 GFLOPS.
- MambaSCI-S achieves 36.31 dB PSNR on Football, over 1.5 dB above the next best method in the large-scale comparison.
- MambaSCI-B requires about 4.5% of STFormer-B's FLOPS (556.89 vs 12155.47 G).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add STMamba (spatial-temporal Mamba) in place of vanilla Mamba (ablation) | MambaSCI Residual-Mamba-Block · 6 simulation color videos | PSNR (dB): +6.13 dB (24.18 to 30.31) | Helped |
| add EDR (edge-detail reconstruction) module (ablation) | MambaSCI with STMamba · 6 simulation color videos | PSNR (dB): +4.35 dB (30.31 to 34.66), but Params 0.28M to 2.53M and FLOPS 35.36G to 235.47G | Helped |
| add CA (channel attention) module (ablation) | MambaSCI with STMamba and EDR · 6 simulation color videos | PSNR (dB): +1.04 dB (34.66 to 35.70), Params 2.53M to 6.11M | Helped |
| remove learnable residual scales (ablation) | Residual-STMamba-Block · 6 simulation color videos | PSNR (dB): -0.37 dB (35.70 to 35.33) | Hurt |
| remove residual connections (ablation) | Residual-STMamba-Block · 6 simulation color videos | PSNR (dB): -0.99 dB (35.70 to 34.71) | Hurt |
| MambaSCI-B vs STFormer-B (quad-Bayer retrained, 6 videos) | MambaSCI · 6 simulation color videos | PSNR (dB): +3.22 dB (35.70 vs 32.48) with 31% params and 4.5% FLOPS | Helped |
| MambaSCI-S vs PnP-FastDVDnet on Messi/Hummingbird | MambaSCI-S · 4 large-scale simulation videos | PSNR (dB): Messi 26.36 vs 29.17 for PnP-FastDVDnet; Hummingbird 30.73 vs 32.11 | Hurt |
| Quad-Bayer retraining of existing E2E methods vs MambaSCI on Bayer-designed models | STFormer and EfficientSCI · 6 simulation color videos | visual color fidelity | Hurt |
Surprising findings
- On large-scale Messi and Hummingbird, MambaSCI-S underperforms some model-based baselines (e.g., PnP-FastDVDnet at 29.17 dB on Messi vs 26.36 dB for MambaSCI-S), which the authors attribute to fast motion and limited capacity.
- At B=32 MambaSCI's PSNR (22.44 dB) is below EfficientSCI (23.24 dB) even though its SSIM is higher.
Limitations
- CA module's Conv3d operations substantially increase parameters and FLOPS, creating a complexity-performance trade-off.
- No real quad-Bayer SCI measurements are available, so evaluation is only on simulated data.
- Performance on fast-motion large-scale videos (Messi, Hummingbird) is weaker than some baselines.
Experimental setup
Datasets: DAVIS2017
Benchmarks: Beauty Bosphorus Jockey Runner ShakeNDry Traffic Messi Hummingbird Swinger Football
Metrics: PSNR SSIM Params (M) FLOPS (G)
Baselines: GAP-TV PnP-FFDNet PnP-FastDVDnet EfficientSCI STFormer
Builds on: Mamba MambaIR U-Mamba STFormer EfficientSCI
Closest papers by meaning
- Detail Matters: Mamba-Inspired Joint Unfolding Network for Snapshot Spectral Compressive Imaging (AAAI 2025 · similarity 0.86)
- Binarized Mamba-Transformer for Lightweight Quad Bayer HybridEVS Demosaicing (CVPR 2025 · similarity 0.85)
- MaIR: A Locality- and Continuity-Preserving Mamba for Image Restoration (CVPR 2025 · similarity 0.84)
- MambaIRv2: Attentive State Space Restoration (CVPR 2025 · similarity 0.84)
- Toward Dynamic Non-Line-of-Sight Imaging with Mamba Enforced Temporal Consistency (NeurIPS 2024 · similarity 0.84)
- LC-Mamba: Local and Continuous Mamba with Shifted Windows for Frame Interpolation (CVPR 2025 · similarity 0.84)
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.