EHETM: High-Quality and Efficient Turbulence Mitigation with Events¶

Conference: CVPR 2026 arXiv: 2603.20708 Code: https://github.com/Xavier667/EHETM Area: Scientific Computing / Event Cameras Keywords: Atmospheric Turbulence Mitigation, Event Camera, Polarity-Weighted Gradient, Event Tube Constraint, Motion Decoupling

TL;DR¶

This paper proposes EHETM, the first method to leverage the microsecond temporal resolution of event cameras to overcome the accuracy–efficiency bottleneck of conventional multi-frame turbulence mitigation (TM). Two key physical phenomena are identified—polarity alternation of turbulence-induced events correlated with image gradients, and spatiotemporally coherent "event tubes" formed by dynamic objects—motivating two complementary modules: a polarity-weighted gradient module and an event tube constraint module. EHETM reduces data overhead by 77.3% and system latency by 89.5%, with particularly substantial gains over the state of the art in dynamic-object scenes.

Background & Motivation¶

Background: Atmospheric and thermal turbulence are the primary degradation sources in long-range imaging, introducing random refractive-index fluctuations that cause geometric tilt and spatially varying blur. Existing methods (e.g., DATUM) rely on multi-frame sequences captured by conventional cameras to recover stable scene content.

Limitations of Prior Work: Multi-frame methods face a fundamental accuracy–efficiency trade-off—more frames yield better restoration but incur higher system latency and data overhead. Furthermore, when dynamic objects are present, multi-frame methods struggle to distinguish turbulence-induced jitter from genuine object motion.

Key Challenge: Turbulence mitigation requires substantial temporal redundancy to average out random jitter, yet the frame rate of conventional cameras limits the amount of information obtainable within an acceptable latency budget.

Key Insight: Event cameras asynchronously record brightness changes at microsecond temporal resolution—the motion information contained in a single event window far exceeds that of a conventional frame—enabling high-quality restoration from very few frames.

Core Idea: Two physical phenomena are identified as restoration priors: (1) polarity alternation of turbulence-induced events is correlated with image gradients, providing scene structure cues; (2) dynamic objects form coherent tubular structures in the event stream, providing motion priors.

Method¶

Overall Architecture¶

Conventional frames (few) + event stream → Polarity-Weighted Gradient Module (scene structure recovery) + Event Tube Constraint Module (dynamic object decoupling) → High-quality restoration output.

Key Designs¶

Phenomenon 1: Polarity Alternation of Turbulence Events and Image Gradients:
- Finding: Brightness changes caused by turbulence generate events with alternating positive/negative polarity at image gradient locations in the clean scene—these events effectively "trace" scene edges and texture structures.
- Polarity-Weighted Gradient Module: Event polarity information is used to weight the estimation of image gradients; positive-polarity events indicate positive gradient directions and negative-polarity events indicate negative gradient directions.
- Design Motivation: Conventional methods require multi-frame averaging to recover sharp edges, whereas event polarity directly supplies structural cues—a single event window can yield structure information comparable to many frames.
Phenomenon 2: "Event Tubes" of Dynamic Objects:
- Finding: Dynamic objects (e.g., vehicles, pedestrians) form continuous tubular structures in the spatiotemporal event stream (events generated by continuous object motion are arranged as tubes in space–time), whereas turbulence-induced events exhibit random, irregular distributions.
- Event Tube Constraint Module: The spatiotemporal coherence of event tubes is exploited to separate object motion from turbulence jitter. Tubular structures are detected → object motion trajectories are extracted → object motion is subtracted from the total displacement → the pure turbulence component is obtained.
- Design Motivation: Conventional multi-frame TM methods assume a static scene—when dynamic objects are present, their motion is misattributed to turbulence. Event tubes provide a natural prior for distinguishing the two.
Real-World Dataset Construction:
- Two event–frame turbulence datasets are constructed: (1) an atmospheric turbulence dataset covering long-range outdoor scenes; (2) a thermal turbulence dataset covering imaging near heat sources.
- Both datasets include static and dynamic scenes.
- Design Motivation: Existing turbulence datasets contain only conventional frames with no event data. Real-world data are critical for reliable evaluation.

Loss & Training¶

Training combines pixel-level reconstruction loss, perceptual loss, and a polarity-consistency constraint.

Key Experimental Results¶

Main Results (Comparison with Multi-Frame TM Methods)¶

Method	Frames	PSNR↑	SSIM↑	Latency	Data Volume
DATUM (50 frames)	50	Baseline	Baseline	100%	100%
DATUM (10 frames)	10	Degraded	Degraded	~20%	~20%
EHETM (2 frames + events)	2	SOTA	SOTA	10.5%	22.7%

Dynamic Scene Comparison¶

Method	Static Scenes	Dynamic Scenes (with Moving Objects)	Notes
DATUM	Good	Severe artifacts	Cannot distinguish object motion from turbulence
EHETM	SOTA	Substantially superior (largest margin)	Event tube constraint effectively decouples motion

Ablation Study¶

Configuration	PSNR	Notes
Conventional frames only (baseline)	Baseline	Poor quality with few frames
+ Polarity-weighted gradient	+Large gain	Structural cues from events
+ Event tube constraint	+Further gain	Dynamic object decoupling
Full EHETM	Best	Two modules are complementary

Efficiency Comparison¶

Metric	EHETM vs. DATUM
Data overhead reduction	77.3%
System latency reduction	89.5%

Key Findings¶

EHETM with only 2 frames + events surpasses DATUM using 50 frames—demonstrating that the temporal resolution advantage of event cameras is fully exploited.
Dynamic scenes constitute the scenario where EHETM achieves the largest margin—conventional methods nearly fail entirely in such scenes, while EHETM handles them effectively via the event tube constraint.
The polarity alternation phenomenon holds for both atmospheric and thermal turbulence, indicating physical generality.
The substantial reductions in data overhead and system latency make real-time long-range imaging practically feasible.

Highlights & Insights¶

First systematic application of event cameras to turbulence mitigation: This work opens a new direction for event-based vision in scientific imaging. The microsecond temporal resolution of event cameras is naturally suited to the high-frequency stochastic nature of turbulence.
Discovery of two physical phenomena: (1) the polarity-alternation–gradient correlation and (2) the spatiotemporal coherence of event tubes. These findings not only guide the method design but also provide important physical priors for subsequent research.
Breakthrough in dynamic scenes: The "static scene assumption" underlying conventional TM methods is fundamentally challenged. The event tube constraint offers an elegant solution to the problem of moving objects under turbulence.
Qualitative efficiency improvement: A 77% reduction in data overhead combined with an 89% reduction in latency represents an order-of-magnitude change rather than incremental progress—transforming real-time turbulence mitigation from infeasible to achievable.

Limitations & Future Work¶

The high hardware cost of event cameras limits practical deployment.
The current event tube detection assumes rigid body motion; handling non-rigid motion (e.g., fluids, deformable objects) remains to be explored.
Robustness under extreme turbulence conditions (e.g., strong convective weather) has not been thoroughly evaluated.
The performance characteristics of event cameras under low-light or extreme high-dynamic-range conditions may affect results.
Combining event-camera-based TM with adaptive optics is a promising future direction.

vs. DATUM/TurbNet and other multi-frame methods: Dependence on large numbers of conventional frames leads to high latency and failure in dynamic scenes. EHETM fundamentally changes the information acquisition paradigm through event cameras.
vs. adaptive optics: Hardware-based solutions entail far higher cost and complexity than software-based approaches. EHETM offers a more lightweight computational alternative.
vs. other event camera applications (optical flow / deblurring): Turbulence mitigation represents a new application domain for event-based vision—the stochastic nature of turbulence is naturally complementary to the high temporal resolution of event cameras.

Rating¶

Novelty: ⭐⭐⭐⭐⭐ A fundamentally new event-driven turbulence mitigation paradigm, together with the discovery of two physical phenomena.
Experimental Thoroughness: ⭐⭐⭐⭐ Real-world dataset construction, comprehensive quantitative and qualitative comparisons, and ablation studies.
Writing Quality: ⭐⭐⭐⭐ Clear description of physical phenomena; method design is well-grounded in physical intuition.
Value: ⭐⭐⭐⭐⭐ Significant contributions to both long-range imaging and event-based vision.