Research map / Video / Video segmentation and tracking
EDCFlow: Exploring Temporally Dense Difference Maps for Event-based 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 EDCFlow, a lightweight event-based optical flow network that fuses temporally dense high-resolution feature-difference motion features with a low-resolution cost volume, and can serve as a refinement module for RAFT-like event methods.
Problem
Event-based optical flow estimation with RAFT-like cost-volume methods suffers from redundant computation for temporally dense cost volumes and poor scalability to higher-resolution flow refinement.
Approach
Replace extra temporally dense cost volumes with multi-scale, attention-weighted temporal feature-difference maps computed at 1/4 resolution after feature warping, and adaptively fuse them with correlation features from a 1/8 cost volume via channel attention.
Key results
- On DSEC, EDCFlow reaches EPE 0.72 and 1PE 10.0, improving EPE by about 9% over E-RAFT and 3% over TMA with 2.5M params and 39 ms runtime on RTX 4090.
- Compared to IDNet-4 with similar accuracy, EDCFlow uses about 0.2x the MACs (247G vs 1200G) and is 68% faster.
- As a refinement module on RAFT-like methods, Ours-8-refine reaches EPE 0.71 on DSEC versus 0.77 for TMA-refine's 0.72 with lower MACs than TMA-refine (352G vs 478G).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add difference motion encoding (ablation) | EDCFlow · DSEC test | EPE: 0.82 -> 0.72 | Helped |
| add correlation motion encoding (ablation) | EDCFlow · DSEC test | EPE: 0.83 -> 0.72 | Helped |
| add channel attention to fuse correlation and difference features (ablation) | EDCFlow · DSEC test | EPE: 0.74 -> 0.72 | Helped |
| add multi-scale attention to the difference layer (ablation) | EDCFlow · DSEC test | EPE: 0.74 -> 0.72 | Helped |
| combine difference sampling scales 1, 2, and 5 instead of using one scale (ablation) | EDCFlow · DSEC test | EPE: 0.77 -> 0.72; 0.74 -> 0.72; 0.79 -> 0.72 | Helped |
| use direct addition instead of depthwise separable 3D convolutions for difference encoding (ablation) | EDCFlow · DSEC test | EPE: 0.72 -> 0.75 | Hurt |
| use channel concatenation and 2D convolution instead of depthwise separable 3D convolutions (ablation) | EDCFlow · DSEC test | EPE: 0.72 -> 0.72 | No clear effect |
| use GRU instead of depthwise separable 3D convolutions for difference encoding (ablation) | EDCFlow · DSEC test | EPE: 0.72 -> 0.73 | Hurt |
Surprising findings
- Adding difference motion features alone without correlation features is worse than the full model, and correlation-only ablation drops accuracy by 15%, yet difference-only features are noted as sensitive to noise and less generalizable.
Limitations
- Accuracy on DSEC is comparable rather than clearly superior to IDNet-4 and E-FlowFormer; MVSEC generalization depends on training data quality.
- Difference features are sensitive to noise, which the paper mitigates only by combining with correlation features.
Experimental setup
Datasets: DSEC MVSEC Blinkflow
Benchmarks: DSEC MVSEC
Metrics: EPE AE 1PE 2PE 3PE % Outlier Param (M) MACs (G) Runtime (ms)
Baselines: E-RAFT TMA IDNet-4 IDNet-8 MultiFlow E-FlowFormer EV-FlowNet DCEIFlow MultiCM STE-FlowNet
Builds on: RAFT E-RAFT TMA IDNet
Closest papers by meaning
- Efficient Meshflow and Optical Flow Estimation from Event Cameras (CVPR 2024 · similarity 0.89)
- Graph Neural Network Combining Event Stream and Periodic Aggregation for Low-Latency Event-based Vision (CVPR 2025 · similarity 0.88)
- FlowTrack: Revisiting Optical Flow for Long-Range Dense Tracking (CVPR 2024 · similarity 0.87)
- MemFlow: Optical Flow Estimation and Prediction with Memory (CVPR 2024 · similarity 0.87)
- StreamFlow: Streamlined Multi-Frame Optical Flow Estimation for Video Sequences (NeurIPS 2024 · similarity 0.87)
- DPFlow: Adaptive Optical Flow Estimation with a Dual-Pyramid Framework (CVPR 2025 · similarity 0.87)
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.