Research map / Video / Video segmentation and tracking
Explicit Depth-Aware Blurry Video Frame Interpolation Guided by Differential Curves
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 DC-BVFI, a blurry video frame interpolation framework that uses differential curves theory with UBNet (lifting frames to 3D point maps) and MPNet (estimating camera-space scene flow) to handle depth variations.
Problem
Blurry video frame interpolation methods based on linear or quadratic motion models overlook depth variations during fast object motion, which change object size and degrade interpolation quality.
Approach
Model frame pixels as differential curves in 3D camera space: lift 2D frames to 3D point maps with a transformer-based 3D reconstruction module, then upgrade 2D optical flow (IFNet) to camera-space scene flow using predicted depth and intrinsics, and warp point maps back to the frame for interpolation.
Key results
- Achieves 34.14 PSNR / 0.942 SSIM average on Adobe, versus 33.79 for DeMFI and IQ-VFI.
- Achieves 31.03 PSNR on RBI, a 1.12 dB improvement over prior state of the art (BiT 29.75 average).
- Reaches 32.08 PSNR on GoPro and 33.35 PSNR on YouTube, reported as state of the art.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DC-BVFI instead of DeMFI | blurry video frame interpolation · Adobe Avg | PSNR: 33.79 -> 34.14 | Helped |
| use DC-BVFI instead of DeMFI | blurry video frame interpolation · Adobe Avg | SSIM: 0.940 -> 0.942 | Helped |
| use DC-BVFI instead of IQ-VFI | blurry video frame interpolation · RBI Avg | PSNR: 29.91 -> 31.03 | Helped |
| use DC-BVFI instead of BiT | blurry video frame interpolation · RBI Avg | SSIM: 0.894 -> 0.913 | Helped |
| use DC-BVFI instead of IQ-VFI | blurry video frame interpolation · YouTube Avg | PSNR: 32.94 -> 33.35 | Helped |
| use DC-BVFI instead of IQ-VFI | blurry video frame interpolation · YouTube Avg | SSIM: 0.928 -> 0.932 | Helped |
| use DC-BVFI instead of IQ-VFI | blurry video frame interpolation · GoPro Avg | PSNR: 31.28 -> 32.08 | Helped |
| use DC-BVFI instead of IQ-VFI | blurry video frame interpolation · GoPro Avg | SSIM: 0.904 -> 0.910 | Helped |
| add scene-flow refinement (ablation) | DC-BVFI · RBI | PSNR: 30.98 -> 31.03 | Helped |
| add discriminative deblur instead of direct frame stacking (ablation) | DC-BVFI · RBI | PSNR: 30.66 -> 31.03 | Helped |
| use a 3D architecture instead of a 2D architecture (ablation) | DC-BVFI · RBI | PSNR: 28.91 -> 31.03 | Helped |
| use scene flow instead of optical flow (ablation) | DC-BVFI · RBI | PSNR: 28.80 -> 31.03 | Helped |
Surprising findings
- Ablation shows the 2D architecture with the same backbone drops to 28.91 PSNR on RBI versus 31.03 for the 3D architecture.
Limitations
- Relies on blurry-frame-based optical flow, which is rough under heavy blur, requiring refinement modules.
- Evaluation uses synthetic blur for most datasets; depth ground truth for MiD is estimated from clear frames rather than measured.
Experimental setup
Datasets: Adobe240 RBI GoPro YouTube blurry video dataset
Benchmarks: Adobe240 RBI GoPro YouTube
Metrics: PSNR SSIM MiD tOF
Baselines: QVI TNTT DAIN BIN ALANET UTI-VFI EQVI ANIB DeMFI BiT JNMR IQ-VFI
Builds on: IFNet DeMFI Dust3R quadratic video interpolation
Closest papers by meaning
- Depth-Aware Video Frame Interpolation (CVPR 2019 · similarity 0.88)
- BiM-VFI: Bidirectional Motion Field-Guided Frame Interpolation for Video with Non-uniform Motions (CVPR 2025 · similarity 0.88)
- Disentangled Motion Modeling for Video Frame Interpolation (AAAI 2025 · similarity 0.88)
- Perception-Oriented Video Frame Interpolation via Asymmetric Blending (CVPR 2024 · similarity 0.87)
- VTinker: Guided Flow Upsampling and Texture Mapping for High-Resolution Video Frame Interpolation (AAAI 2026 · similarity 0.86)
- Sparse Global Matching for Video Frame Interpolation with Large Motion (CVPR 2024 · 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.