Research map / Probabilistic and scientific modeling / Molecules and proteins
CEDFlow: Latent Contour Enhancement for Dark Optical Flow Estimation
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 CEDFlow, a RAFT-style optical flow network for dark images with a dual-branch frequency-decomposed motion encoder and a latent-space 2nd-order Gaussian difference contour enhancement, reporting lower EPE on FCDN and VBOF.
Problem
Optical flow estimation in low-contrast, noisy dark images where contour cues are degraded and standard flow networks that rely on high-contrast texture fail.
Approach
Decompose features into high- and low-frequency parts via learned contrast-aware attention, encode them with separate local and global encoders, then select and enhance large-scale contours directly in latent feature space using a 2nd-order Gaussian difference (D2oG) weighting with double-pass sigmoid filtering.
Key results
- CEDFlow reaches mean EPE 1.08 on FCDN when trained on FCDN, vs 1.15 for the second-best AGFlow.
- Trained on mixed data, CEDFlow achieves EPE 4.69 on VBOF Fuji2 and 6.52 on VBOF All, the best among the compared methods.
- Removing the latent contour enhancement module raises FCDN EPE from 1.08 to 1.26 in the ablation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CEDFlow instead of AGFlow | Optical flow estimation · FCDN (trained on FCDN) | EPE: 1.15 →1.08 | Helped |
| Use CEDFlow instead of AGFlow | Optical flow estimation · VBOF (Fuji2, trained on FCDN) | EPE: 14.16 -> 13.94 | Helped |
| Use CEDFlow instead of AGFlow | Optical flow estimation · VBOF (All, trained on FCDN) | EPE: 21.05 -> 20.89 | Helped |
| Use CEDFlow instead of GMA | Optical flow estimation · FCDN (trained on Mixed) | EPE: 1.26 -> 1.23 | Helped |
| Use CEDFlow instead of GMA | Optical flow estimation · VBOF (Fuji2, trained on Mixed) | EPE: 4.90 -> 4.69 | Helped |
| Use CEDFlow instead of AGFlow | Optical flow estimation · VBOF (All, trained on Mixed) | EPE: 6.75 -> 6.52 | Helped |
| Set Gaussian radius r to 3 rather than 1, 2, or 4 (ablation) | CEDFlow D2oG contour selection · FCDN and VBOF (Fuji2) | EPE: FCDN: 1.27, 1.17, 1.08, 1.21; VBOF(Fuji2): 14.22, 13.99, 13.94, 14.15 | Helped |
| Use Gaussian scales σn=(3, 9, 27) rather than (1, 3, 9) or (2, 6, 18) (ablation) | CEDFlow D2oG contour selection · FCDN and VBOF (Fuji2) | EPE: FCDN: 1.17, 1.18, 1.08; VBOF(Fuji2): 14.14, 14.33, 13.94 | Helped |
| Replace RAFT encoder with DBME (ablation) | RAFT · FCDN and VBOF (Fuji2), trained on FCDN | EPE: FCDN: 1.23 -> 1.14; VBOF(Fuji2): 14.20 -> 14.13 | Helped |
| Add LCE (ablation) | RAFT · FCDN and VBOF (Fuji2), trained on FCDN | EPE: FCDN: 1.23 -> 1.12; VBOF(Fuji2): 14.20 -> 14.03 | Helped |
| Replace DBME with RAFT encoder (ablation) | CEDFlow · FCDN and VBOF (Fuji2), trained on FCDN | EPE: FCDN: 1.08 -> 1.17; VBOF(Fuji2): 13.94 -> 14.38 | Hurt |
| Add feature decomposition (ablation) | CEDFlow · FCDN, VBOF (Fuji2), and VBOF (All), trained on FCDN | EPE: FCDN: 1.19 -> 1.08; VBOF(Fuji2): 14.70 -> 13.94; VBOF(All): 23.13 -> 20.89 | Helped |
Limitations
- Gains on VBOF Fuji2 under FCDN training are modest relative to some baselines and the parameter count is second largest among compared models.
- Gaussian kernel parameters (r, sigma) are sensitive and tuned per dataset.
Experimental setup
Datasets: FCDN VBOF
Benchmarks: FCDN VBOF FLIR ADAS
Metrics: EPE parameter count runtime (ms) memory (GB)
Baselines: RAFT GMFlowNet GMFlow AGFlow GMA KPAFlow SCV Flow1D
Builds on: RAFT GMA FlowNet
Closest papers by meaning
- Shape and Texture: What Influences Reliable Optical Flow Estimation? (CVPR 2025 · similarity 0.85)
- DPFlow: Adaptive Optical Flow Estimation with a Dual-Pyramid Framework (CVPR 2025 · similarity 0.85)
- Exploring the Common Appearance-Boundary Adaptation for Nighttime Optical Flow (ICLR 2024 · similarity 0.85)
- Recurrent Partial Kernel Network for Efficient Optical Flow Estimation (AAAI 2024 · similarity 0.84)
- FlowDiffuser: Advancing Optical Flow Estimation with Diffusion Models (CVPR 2024 · similarity 0.83)
- Multi-Modal Synergistic Implicit Image Enhancement for Efficient Optical Flow Estimation (CVPR 2025 · 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.