Skip to content

Low-Light Image Enhancement using Event-Based Illumination Estimation (RetinEV)

Conference: ICCV 2025 arXiv: 2504.09379 Code: Not released Area: Low-Light Image Enhancement / Event Camera Keywords: Low-light enhancement, event camera, Retinex theory, temporal-mapping events, illumination estimation, reflectance enhancement

TL;DR

RetinEV proposes exploiting temporal-mapping events (triggered by transmittance modulation) rather than conventional motion events for illumination estimation. Combined with Retinex theory, it decomposes low-light images into illumination and reflectance components, and employs an Illumination-guided Reflectance Enhancement (IRE) module to achieve high-quality low-light image enhancement, reaching real-time inference at 35.6 FPS on 640×480 images.

Background & Motivation

Background: Low-light image enhancement (LLIE) is a fundamental task in computer vision. Traditional approaches fall into two categories: histogram equalization and Retinex-based methods. Deep learning methods (RetinexNet, SCI, Diff-Retinex, etc.) have achieved notable progress. Event cameras have attracted attention due to their high dynamic range and superior low-light response.

Limitations of Prior Work: - Existing event-guided LLIE methods (e.g., eSL-Net, EvLight) rely on motion events (triggered by object or camera motion), which suffer from three key problems: - (a) Low-light conditions require long exposure, during which motion events compound motion blur degradation. - (b) Motion events provide only edge information, failing to cover flat regions for illumination estimation. - (c) They are prone to generating artifacts. - Image-only methods lack effective illumination priors under extreme low-light conditions.

Key Challenge: Event cameras possess excellent HDR and low-light response capabilities, yet existing methods exploit only motion-triggered events, leaving the potential of event cameras along the illumination dimension largely untapped.

Key Insight: A paradigm shift—rather than relying on scene motion, the paper actively generates temporal-mapping events by varying the optical transmittance (opening a mechanical shutter). These events directly encode scene brightness changes, providing denser and more accurate illumination information than motion events. Converting event timestamps to brightness values yields fine-grained illumination estimation.

Method

Overall Architecture

RetinEV is built on Retinex theory: observed image = reflectance × illumination (\(I = R \cdot S\)). The pipeline consists of:

  1. Illumination Estimation: Estimate the illumination component \(S\) from temporal-mapping events.
  2. Image Decomposition: A decomposition network \(\mathcal{F}_{Decom}\) decomposes both low-light and normal-light images into reflectance \(R\) and illumination \(S\).
  3. Reflectance Enhancement: The IRE module uses estimated illumination as a prior to enhance detail in the reflectance component.
  4. Brightness Control: A coefficient \(\beta\) enables arbitrary control of output brightness.

During training, normal-light images serve as supervision, and reflectance consistency constraints ensure that low-light and normal-light images share the same reflectance.

Key Designs

  1. Temporal-Mapping Event-Driven Illumination Estimation:

    • Function: Opening the mechanical shutter (taking ~2 ms) transitions the optical transmittance from 0 to 1, triggering dense event signals on the event camera.
    • Mechanism: For pixel \((x,y)\), the event timestamp \(t(x,y)\) is inversely proportional to pixel brightness \(I(x,y)\)—brighter pixels trigger events earlier due to faster brightness change, while darker pixels trigger later. Converting event timestamps to brightness values thus yields precise illumination estimates.
    • Design Motivation: Motion events have limited information density (triggered only at motion edges), whereas temporal-mapping events provide dense illumination information across the entire image, independent of scene motion.

  2. Illumination-guided Reflectance Enhancement (IRE) Module:

    • Function: Uses the estimated illumination component as a prior and enhances the reflectance component via cross-modal attention.
    • Mechanism: Conventional Retinex methods focus solely on enhancing the illumination component, but under low light the reflectance is also severely affected by noise and color bias. IRE leverages high-quality illumination priors to guide reflectance denoising and detail recovery.
    • Design Motivation: The illumination component encodes the scene's overall brightness structure, which helps identify noise patterns and textural details within the reflectance.

  3. Degradation Model for Temporal-Mapping Events under Low Light:

    • Function: Models the effect of low-light conditions on event camera behavior, specifically changes in the event triggering threshold.
    • Mechanism: In low-light scenes, the SNR of event cameras decreases, and events in dark regions may be missing or noisy. The degradation model brings synthetic training data closer to real low-light conditions, improving generalization.

  4. Brightness Controllability: A coefficient \(\beta\) multiplied by the illumination component \(S\) allows users to freely adjust the brightness of the output image.

Hardware & Dataset

  • Beam-splitter System: A coaxial optical system combining a DVS event camera and an RGB image sensor, ensuring precise alignment between the event stream and the low-light image.
  • EvLowLight Dataset: 60 extremely low-illumination and high-contrast scenes, containing aligned images, temporal-mapping events, and motion events.
  • The mechanical shutter opening process (~2 ms) simultaneously accomplishes image exposure and temporal-mapping event acquisition.

Loss & Training

  • Decomposition loss: enforces reflectance consistency between low-light and normal-light images.
  • Reconstruction loss: constrains the quality of enhanced images.
  • Event degradation augmentation: accounts for low-light event camera characteristics on synthetic data.
  • Training data: 5 synthetic datasets + the EvLowLight real-world dataset.

Key Experimental Results

Main Results

Comprehensive evaluation on 5 synthetic datasets and the EvLowLight real-world dataset:

  • vs. motion-event methods: RetinEV surpasses the previous state-of-the-art event-based method (EvLight) by up to 6.62 dB PSNR.
  • vs. image-only methods: Consistently outperforms traditional LLIE methods on both reference-based and no-reference objective metrics, as well as subjective user studies.
  • Inference speed: Achieves real-time processing at 35.6 FPS on 640×480 images.

Core Advantages over Motion-Event Methods

Dimension Motion-Event Methods RetinEV (Temporal-Mapping Events)
Event source Scene/camera motion Transmittance modulation (shutter opening)
Information density Edge regions only Full image coverage
Illumination information Indirect (requires inference) Directly encodes brightness
Motion blur Severe under low light Independent of motion; not an issue
Artifacts Prone to generation Significantly reduced

Key Findings

  • Temporal-mapping events are a superior event signal for LLIE: From an information-theoretic perspective, temporal-mapping events directly encode illumination information, making them more suitable for illumination estimation than the indirect edge information provided by motion events.
  • Effectiveness of the IRE module: Jointly enhancing both reflectance and illumination outperforms traditional Retinex methods that enhance only the illumination component.
  • Importance of event degradation modeling: Modeling the impact of low light on event cameras leads to significantly improved generalization from synthetic training data to real-world data.
  • Lightweight and efficient architecture: Compared to generative model approaches (GAN/diffusion models), RetinEV is more lightweight, achieving real-time inference at 35.6 FPS.

Highlights & Insights

  • Value of the paradigm shift: The transition from "motion events" to "temporal-mapping events" is a remarkably clever paradigm shift. By actively generating events (modulating transmittance) rather than passively waiting for them (scene motion), the HDR and low-light capabilities of event cameras are transformed from "potentially useful" to "directly applicable for illumination estimation."
  • Deep exploitation of physical priors: The conversion from event timestamps to brightness values has a clear physical basis (the correspondence between transmittance change and brightness), making this physics-prior-based design more interpretable and generalizable than purely data-driven approaches.
  • Elegant cross-modal fusion design: The IRE module injects illumination priors into reflectance enhancement, representing an insightful cross-modal attention design. It recognizes that illumination and reflectance in Retinex decomposition are not independent—high-quality illumination estimates can in turn assist reflectance denoising.
  • Complete system solution: From hardware (beam-splitter system) to dataset (EvLowLight) to algorithm (RetinEV), the paper provides a comprehensive end-to-end solution that facilitates follow-up research.

Limitations & Future Work

  • Hardware dependency: Requires a dedicated beam-splitter system integrating both DVS and RGB sensors, as well as a precisely controlled mechanical shutter, limiting deployment convenience.
  • Acquisition constraints for temporal-mapping events: Events generated by shutter opening must be captured at the start of exposure, imposing special requirements on the capture workflow and making the approach unsuitable for continuous video enhancement.
  • Limited scale of EvLowLight dataset: With only 60 scenes, the dataset may not cover a sufficiently diverse range of low-light degradation patterns.
  • Extreme low-light scenarios near the event noise floor: Event cameras may fail to trigger events effectively in very dark regions, leading to incomplete illumination estimates.
  • Incomplete comparison with latest diffusion-based LLIE methods: Methods such as DiffLL may offer advantages in perceptual quality, though their inference speed is far below RetinEV's real-time performance.

Highlights & Insights

Limitations & Future Work

Rating

  • Novelty: Pending
  • Experimental Thoroughness: Pending
  • Writing Quality: Pending
  • Value: Pending