Research map / Video / Video segmentation and tracking
GaraMoSt: Parallel Multi-Granularity Motion and Structural Modeling for Efficient Multi-Frame Interpolation in DSA Images
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 GaraMoSt, a parallel multi-granularity motion-structure feature extractor (MG-MSFE) for direct multi-frame interpolation of DSA angiography images, improving noise suppression over MoSt-DSA at similar inference time.
Problem
Direct multi-frame interpolation of DSA images needs accurate, real-time output; existing natural-scene VFI methods and MoSt-DSA leave high- and low-frequency noise, motion artifacts and blurring in interpolated frames.
Approach
A parallel network pipeline with two feature paths at different granularities, where Lambda-layer-style linear context interactions (instead of attention maps) extract relative motion and structural features at multiple scales in a fully convolutional way, followed by dual-layer flow-mask estimation and refinement.
Key results
- Single-frame: GaraMoSt-L1 reaches 95.05 mean SSIM and 40.69 PSNR vs. MoSt-DSA-L1 94.62 and 40.32, at 0.029 s vs 0.024 s.
- Two-frame: GaraMoSt achieves 94.01 mean SSIM (+0.87 over MoSt-DSA) and 39.66 PSNR.
- Three-frame: GaraMoSt-L1 reaches 94.22 mean SSIM (+0.64 over MoSt-DSA) with inference time 0.122 s.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GaraMoSt-L1 instead of MoSt-DSA-L1 | DSA interpolation · Single-frame interpolation | SSIM mean (%): 94.62 -> 95.05 | Helped |
| Use GaraMoSt instead of MoSt-DSA | DSA interpolation · Two-frame interpolation | SSIM mean (%): 93.14 -> 94.01 | Helped |
| Use GaraMoSt-L1 instead of MoSt-DSA-L1 | DSA interpolation · Three-frame interpolation | SSIM mean (%): 93.58 -> 94.22 | Helped |
| Use granularity 7, 29 rather than 7, 7 (ablation) | GaraMoSt · Three-frame interpolation | SSIM mean (%): 93.41 -> 93.49 | Helped |
| Use granularity 29, 15 rather than 7, 7 (ablation) | GaraMoSt · Single-frame interpolation | SSIM mean (%): 94.37 -> 94.18 | Hurt |
| Use granularity 15, 15 rather than 7, 7 (ablation) | GaraMoSt · Single-frame interpolation | SSIM mean (%): 94.37 -> 94.31 | Hurt |
| Replace shallow refiner features with upsampled deep structural features (ablation) | GaraMoSt · Single-frame interpolation | SSIM mean (%): 94.37 -> 94.27 | Hurt |
| Replace shallow refiner features with upsampled deep structural features (ablation) | GaraMoSt · Single-frame interpolation | PSNR mean: 40.19 -> 40.10 | Hurt |
Surprising findings
- Using shallow structural features in the refiner outperforms deeper ones; replacing them with upsampled deeper features lowers mean SSIM by 0.10.
Limitations
- Gains over baselines in SSIM/PSNR are small in absolute terms, and the authors note minimal score differences among SOTA VFI methods.
- Evaluated only on a single private DSA dataset of head sequences at 489x489 resolution.
Experimental setup
Datasets: MoSt-DSA DSA dataset (470 head DSA sequences)
Benchmarks: MoSt-DSA DSA dataset (single, two, three-frame interpolation) UCF101
Metrics: SSIM mean SSIM STD PSNR mean PSNR STD inference time
Baselines: MoSt-DSA EMA-VFI EMA-small FILM ABME SoftSplat
Builds on: MoSt-DSA Lambda Layer EMA-VFI FILM SoftSplat
Closest papers by meaning
- VasTSD: Learning 3D Vascular Tree-state Space Diffusion Model for Angiography Synthesis (CVPR 2025 · similarity 0.82)
- CompactSplat: Spatially Adaptive Gaussian Distribution for Feedforward Scene Reconstruction (NeurIPS 2026 · similarity 0.81)
- DGS-LRM: Real-Time Deformable 3D Gaussian Reconstruction From Monocular Videos (NeurIPS 2025 · similarity 0.81)
- MoDec-GS: Global-to-Local Motion Decomposition and Temporal Interval Adjustment for Compact Dynamic 3D Gaussian Splatting (CVPR 2025 · similarity 0.81)
- Blood Flow Speed Estimation with Optical Coherence Tomography Angiography Images (CVPR 2025 · similarity 0.81)
- VFIMamba: Video Frame Interpolation with State Space Models (NeurIPS 2024 · similarity 0.81)
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.