Research map / Video / Video segmentation and tracking
RivuletMLP: An MLP-based Architecture for Efficient Compressed Video Quality Enhancement
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 RivuletMLP, an MLP-based network for compressed video quality enhancement that combines deformable multi-frame alignment, permutation-based global feature flow, and local-refinement compensation, reaching higher PSNR gain than STCF at lower FLOPs on MFQEv2.0.
Problem
Compressed video quality enhancement needs long-range spatiotemporal modeling and aligned multi-frame information, but CNNs have limited receptive fields and Transformers are computationally expensive.
Approach
Replaces attention with patch-based MLP layers plus a feature permutation (flow) mechanism that moves features from distant positions into local MLP windows for non-local dependencies, combined with a dynamically guided deformable alignment module and a compensation module for inter-frame motion.
Key results
- At QP=37 on MFQEv2.0 (LDP), RivuletMLP achieves average ΔPSNR 1.13 dB vs 1.02 dB for STCF and 0.92 dB for CPGA.
- RivuletMLP uses 615.7 GFLOPs versus 1677.1 GFLOPs for STCF at 720p, roughly 36.6% of STCF's FLOPs.
- The lightweight variant (c=64) reaches 0.90 dB ΔPSNR at 178.5 GFLOPs and 23.2 FPS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the full RivuletMLP architecture | RivuletMLP · MFQEv2.0 LDP QP=37 average | ∆PSNR (dB): 1.02 -> 1.13 | Helped |
| Use the full RivuletMLP architecture | RivuletMLP · MFQEv2.0 LDP QP=37 average | ∆SSIM (×10−2): 1.82 -> 1.93 | Helped |
| Use the full RivuletMLP architecture | RivuletMLP · MFQEv2.0 LDP QP=42 average | ∆PSNR (dB): 0.99 -> 1.01 | Helped |
| Use the full RivuletMLP architecture | RivuletMLP · MFQEv2.0 LDP QP=32 average | ∆PSNR (dB): 1.07 -> 1.05 | Hurt |
| Use the full RivuletMLP architecture | RivuletMLP · MFQEv2.0 LDP QP=27 average | ∆PSNR (dB): 1.05 -> 0.95 | Hurt |
| Add SFF to BSC (ablation) | RivuletMLP · MFQEv2.0 LDP | ∆PSNR (dB): 0.71 -> 1.13 | Helped |
| Add BSC to SFF (ablation) | RivuletMLP · MFQEv2.0 LDP | ∆PSNR (dB): 0.95 -> 1.13 | Helped |
| Reduce feature dimension from 128 to 64 (ablation) | RivuletMLP · MFQEv2.0 LDP | ∆PSNR (dB): 1.13 -> 0.90 | Hurt |
| Reduce module count from 12 to 8 (ablation) | RivuletMLP · MFQEv2.0 LDP | ∆PSNR (dB): 1.13 -> 1.04 | Hurt |
| Replace STDF alignment with DDA alignment (ablation) | RivuletMLP · MFQEv2.0 LDP | ∆PSNR (dB): 1.07 -> 1.13 | Helped |
Surprising findings
- The MLP-based model has more parameters (6.9M) than STCF (7.0M is listed, paper says 2% less) yet is far cheaper in FLOPs, showing FLOPs and parameters do not track together.
Limitations
- Parameter count is higher than the lightweight baselines in some comparisons
- CPGA had to be retrained by the authors, and some baseline numbers are cited from papers
- Evaluated only on MFQEv2.0 with HEVC LDP configuration
Experimental setup
Datasets: MFQEv2.0
Benchmarks: MFQEv2.0
Metrics: PSNR SSIM delta PSNR delta SSIM FLOPs FPS Params
Baselines: MFQEv2.0 STDF RFDA TVQE STCF CPGA
Builds on: STDF MLP-Mixer SKFF Vision Permutator
Closest papers by meaning
- Multi-Frame Deformable Look-Up Table for Compressed Video Quality Enhancement (AAAI 2025 · similarity 0.85)
- FMA-Net: Flow-Guided Dynamic Filtering and Iterative Feature Refinement with Multi-Attention for Joint Video Super-Resolution and Deblurring (CVPR 2024 · similarity 0.84)
- Plug-and-Play Versatile Compressed Video Enhancement (CVPR 2025 · similarity 0.84)
- FPSAttention: Training-Aware FP8 and Sparsity Co-Design for Fast Video Diffusion (NeurIPS 2025 · similarity 0.84)
- CPGA: Coding Priors-Guided Aggregation Network for Compressed Video Quality Enhancement (CVPR 2024 · similarity 0.83)
- Blur-aware Spatio-temporal Sparse Transformer for Video Deblurring (CVPR 2024 · similarity 0.83)
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.