Research map / Detection and segmentation / CNN features
One-Step Event-Driven High-Speed Autofocus
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 Event Laplacian Product (ELP), a focus detection function combining event frames with grayscale Laplacians, enabling one-step event-driven autofocus that detects the focus position's sign change in real time.
Problem
Event-driven autofocus methods still require capturing a complete focus stack (a lengthy round of focus hunting) before locating the focus position, which is slow and inefficient.
Approach
Derives a link between the spatial second-order derivative of the image and the temporal first-order derivative during focusing, yielding the ELP function whose sign flips from positive to negative at focus; this enables real-time one-step focusing that also indicates motion direction, combined with event-based temporal-mapping photography for event-only cameras.
Key results
- ELP reduces MAE to 0.29 µm on DAVIS346 vs 7.02 µm for PBF (about 24x lower) and 0.57 µm vs 12.54 µm on EVK4 (about 22x lower).
- ELP focusing time is about 103 ms on DAVIS346 and 113 ms on EVK4 versus about 302-310 ms for EGS and PBF, roughly one-third.
- On synthetic data under static, moderate and violent motion, ELP (50 FPS) achieves MAE of 2.00, 1.51 and 2.26 µm versus 4.93, 3.99, 2.69 µm for PBF.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ELP instead of PBF | Event-driven autofocus · DAVIS346 real datasets | MAE (µm): 7.02 -> 0.29 | Helped |
| Use ELP instead of PBF | Event-driven autofocus · EVK4 real datasets | MAE (µm): 12.54 -> 0.57 | Helped |
| Use ELP instead of EGS | Event-driven autofocus · DAVIS346 real datasets | Focusing time (ms): 310.81 -> 103.36 | Helped |
| Use ELP instead of EGS | Event-driven autofocus · EVK4 real datasets | Focusing time (ms): 404.25 -> 112.97 | Helped |
| Increase grayscale frame rate from 1 FPS to 50 FPS | ELP · Synthetic datasets, moderate motion | MAE (µm): 3.66 -> 1.51 | Helped |
| Remove the adaptive ELP filter (ablation) | ELP · DAVIS346 real datasets | MAE (µm): 0.29 -> 2.12 | Hurt |
| Remove the adaptive ELP filter (ablation) | ELP · EVK4 real datasets | MAE (µm): 0.57 -> 19.85 | Hurt |
| Add the Laplacian instead of using a uniform 1 matrix (ablation) | ELP · DAVIS346 real datasets | MAE (µm): 2.48 -> 0.29 | Helped |
| Add the Laplacian instead of using a uniform 1 matrix (ablation) | ELP · EVK4 real datasets | MAE (µm): 7.81 -> 0.57 | Helped |
| Remove Laplacian information (ablation) | ELP · DAVIS346 real datasets | focus adjustment direction | Hurt |
Surprising findings
- Without the Laplacian term, ELP degenerates to the difference of negative and positive event rates and gives the wrong focus direction in 40% of DAVIS cases.
- EGS, the earlier event-rate method, fails to produce a focus position in 28.6% of synthetic cases.
Limitations
- Faces challenges under extreme high-speed motion in event-only settings.
- Relies on a grayscale image (or EvTemMap-derived grayscale) for the Laplacian; fluctuations from noisy short event frames require an adaptive filter.
Experimental setup
Datasets: DAVIS346 real focus stacks Prophesee EVK4 real focus stacks synthetic PSF-based focus events (V2E simulated)
Benchmarks: synthetic PSF focus event dataset DAVIS346 dataset Prophesee EVK4 dataset
Metrics: MAE (µm) runtime (ms) focusing time (ms)
Baselines: EGS PBF PDAF
Builds on: PBF polarity-based autofocus EGS event golden search EvTemMap event-driven temporal-mapping photography V2E
Closest papers by meaning
- EvFocus: Learning to Reconstruct Sharp Images from Out-of-Focus Event Streams (ICML 2025 · similarity 0.86)
- Intensity-Robust Autofocus for Spike Camera (CVPR 2024 · similarity 0.84)
- Event Fields: Capturing Light Fields at High Speed, Resolution, and Dynamic Range (CVPR 2025 · similarity 0.84)
- Stabilizing and Accelerating Autofocus with Expert Trajectory Regularized Deep Reinforcement Learning (CVPR 2025 · similarity 0.83)
- Towards Robust Event-guided Low-Light Image Enhancement: A Large-Scale Real-World Event-Image Dataset and Novel Approach (CVPR 2024 · similarity 0.83)
- EF-3DGS: Event-Aided Free-Trajectory 3D Gaussian Splatting (NeurIPS 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.