EGTM: Synchronized Event-Frame Turbulence Dataset

• EGTM dataset is the first real-world synchronized event–RGB turbulence benchmark, pairing event streams with RGB frames to test event-guided turbulence mitigation.
• It employs a dedicated acquisition system with a heat chamber, co-optical-axis sensors, and microsecond-level synchronization for precise spatial-temporal alignment.
• The benchmark comprises 793 samples with controlled turbulent distortions, supporting efficient evaluation and ablation studies that highlight significant PSNR and SSIM improvements.

The EGTM dataset is the dataset introduced in "EGTM: Event-guided Efficient Turbulence Mitigation" as the first real-world synchronized event–RGB turbulence benchmark designed specifically for event-guided turbulence mitigation (TM) (Li et al., 4 Sep 2025). It was created because existing turbulence mitigation datasets are frame-only and therefore unsuitable for methods that use synchronized event streams to guide restoration. In the paper’s formulation, the dataset is not an auxiliary resource but the empirical basis for testing the central "event-lucky insight": an inverse relationship between event density and image quality, in which more events often correspond to stronger turbulent distortion and fewer events indicate more stable, "lucky" regions.

1. Origin and problem setting

The dataset addresses a specific limitation in prior TM benchmarks: they contain RGB or frame sequences degraded by atmospheric turbulence, but no synchronized event data (Li et al., 4 Sep 2025). That omission matters because conventional multi-frame TM methods must infer turbulence cues implicitly from temporally sparse RGB frames, whereas the EGTM framework is explicitly built around event streams.

The paper motivates the dataset through the contrast between frame cameras and event cameras. Frame-only methods attempt to recover turbulence-free content from synchronous frames with limited frame-rate, which the paper characterizes as computationally heavy and inefficient. Event cameras instead capture microsecond-level asynchronous brightness changes, and the dataset was constructed to test whether those measurements expose turbulence dynamics more directly. The core hypothesis is the paper’s "event-lucky insight": turbulent distortion and the inverse spatiotemporal distribution of event streams are correlated, so event activity can serve as a cue for locating turbulence-free regions.

This design implies a narrower but technically precise target domain. The dataset is intended for event-guided turbulence mitigation, not for generic motion deblurring or arbitrary multimodal restoration. A plausible implication is that its most distinctive contribution lies in enabling learning and evaluation under a sensing configuration that previous TM datasets did not provide.

2. Data acquisition system and sensing configuration

The authors built a dedicated turbulence data acquisition system consisting of three components: a heat chamber / turbulence generator, a hybrid co-optical-axis imaging system, and a screen for controlled target display (Li et al., 4 Sep 2025). The turbulence is generated using a heat chamber system placed 15 meters from the target, producing temperature gradients of 20–60°C. The paper states that this follows prior turbulence protocols and is intended to simulate real atmospheric turbulence in a controlled way.

Scene presentation is also controlled. A screen displays the target image, and ground truth images are captured when the heat source is disabled. Test scenes are displayed automatically via a computer-controlled screen. The setup is therefore a real-world controlled turbulence dataset, but not an outdoor long-range natural-turbulence benchmark.

The imaging system is co-optical-axis. A beam splitter divides incoming light so that both sensors observe the same scene simultaneously. The event sensor is Prophesee EVK5, and the frame camera is FLIR Grasshopper3. The beam splitter is the Thorlabs BSW26R. The event stream is represented as

$\mathcal{E} = \{(x_i, y_i, t_i, p_i)\},$

where $x_i, y_i$ are pixel coordinates, $t_i$ is the timestamp, and $p_i$ is polarity.

The paper emphasizes two alignment properties. Temporal synchronization uses programmable trigger circuits with microsecond precision, specifically

$< 10\mu s,$

and spatial alignment is achieved by ROI cropping and stereo rectification. These steps are necessary because the dataset is intended for pixel-level multimodal learning rather than loose cross-modal correspondence.

3. Dataset composition and benchmark structure

The real-world EGTM dataset contains 793 samples sourced from MIT Places (Li et al., 4 Sep 2025). Each sample includes 36 turbulent frames, synchronized event streams, and one ground-truth reference image. The cropped resolution is

$256 \times 256,$

the frame rate is

$20\ \text{FPS},$

and the sequence duration is

$1.8\ \text{s}.$

The train/test split is 634 samples for training and 159 samples for testing. The split is explicitly scene-level, so scenes are separated across training and testing. This is meant to prevent overfitting and leakage.

The paper also specifies the temporal configuration used in experiments. The model uses 11 frames as input during training and evaluation. In the event representation, the EDEM converts event streams into voxels with

$B = 4N,$

and since $N=11$,

$x_i, y_i$0

The voxelized representation is written as

$x_i, y_i$1

A synthetic event-driven TM benchmark is also mentioned, created using the v2e event simulator on ATSyn-static, but the real-world EGTM dataset is the benchmark identified as the core contribution. This suggests that the paper treats synthetic data as supplementary, while the principal novelty lies in the real synchronized event–frame acquisition.

4. Capture conditions, annotation, and preprocessing

The dataset is collected in an indoor / controlled experimental setting with static or quasi-static scenes, and the paper states that turbulence dominates temporal variations (Li et al., 4 Sep 2025). Because the scene content is displayed on a screen, there is no significant object motion. This is important for interpretation of the event stream: the event-lucky hypothesis assumes that events arise mainly from turbulence-induced brightness changes rather than from scene motion.

The severity of turbulence is controlled through the thermal gradients of

$x_i, y_i$2

with the target at

$x_i, y_i$3

These conditions are central to the benchmark’s intended use. They isolate turbulence-dominated dynamics rather than blending turbulence with independently moving foreground objects.

For supervision, each sample has a ground-truth reference image captured with the heat source turned off. Spatial alignment uses ROI cropping and stereo rectification, and synchronization uses programmable trigger circuits with precision better than 10 microseconds. The dataset is validated by spatial-temporal consistency checks and manual inspection.

Preprocessing is modest but explicit. Images are cropped to $x_i, y_i$4, and event streams are voxelized by time-slice accumulation. The paper’s qualitative descriptions further characterize the collected data: turbulent events appear densely in regions with strong intensity fluctuations; the real-world data exhibit spatial distortion, random blur, temporal instability, and regions of varying event density corresponding to "lucky" and "unlucky" areas.

5. Use in EGTM and empirical findings

The dataset is central to the EGTM method because EGTM uses event streams to estimate pixel-wise lucky fusion weights (Li et al., 4 Sep 2025). The temporal weighting is written as

$x_i, y_i$5

and the fused image is

$x_i, y_i$6

The final refinement stage is

$x_i, y_i$7

and the training objective is

$x_i, y_i$8

with

$x_i, y_i$9

The dataset is also used to validate the event-lucky insight directly. The paper samples 100 randomly selected samples and 500 uniformly sampled pixels per sample, then measures the correlation between local event density and reconstruction error $t_i$0 between turbulent and ground-truth pixels. Spatial windows range from

$t_i$1

and temporal windows from

$t_i$2

The reported Pearson correlations are

$t_i$3

which the paper interprets as evidence that higher event density correlates with larger degradation.

The dataset supports baseline comparison against TurbNet, ATnet, TSRWGAN, VRT, RNN-MBP, ESTRNN, RVRT, and DATUM. On the real-world EGTM dataset, EGTM achieves PSNR 34.38 and SSIM 0.9339, while the strongest competing baseline is RNN-MBP with 33.44 PSNR / 0.9262 SSIM. The abstract states improvements of $t_i$4 PSNR and $t_i$5 SSIM. The details also note that the table values imply an SSIM improvement of 0.0077 in raw terms, and explicitly describe this as a likely rounded or formatting inconsistency in the manuscript text. That inconsistency is part of the record and is best read as a reporting issue rather than a conceptual one.

Efficiency claims are also tied to this benchmark. The paper reports 0.02M parameters, 1.5G FLOPs, and 2 ms latency for EGTM, and claims 214× faster inference and 224× smaller model. The abstract additionally states 710 times, 214 times, and 224 times in model size, inference latency, and model complexity respectively. The dataset thus functions not only as a quality benchmark but also as a testbed for the paper’s efficiency argument.

An ablation study on the real-world dataset reports the following progression:

Configuration	PSNR	SSIM
Inverse Voxel	26.73	0.7451
+SGEB	28.52	0.7914
+TGEB	29.76	0.8038
+SGEB+TGEB	32.25	0.9203
All Components	34.38	0.9339

The paper uses these results to argue that raw inverse event weighting is insufficient, that both spatial and temporal guidance matter, and that the detail extractor provides an additional boost.

6. Scope, limitations, and significance

The paper frames the dataset as the first synchronized event-frame turbulence dataset for event-guided turbulence mitigation and therefore as the enabling condition for this line of work (Li et al., 4 Sep 2025). Its significance follows from the fact that prior TM benchmarks were frame-only. By supplying synchronized event streams, turbulent RGB frames, and ground-truth reference images under controlled turbulence, it allows direct study of whether event activity can guide lucky fusion more efficiently than conventional frame-only inference.

At the same time, the dataset has clear scope conditions. It is built for static/quasi-static, turbulence-dominated scenes, so it does not directly address complex moving objects. The paper also notes that turbulent events are noisy, which is why learning-based guidance is needed, and that synthetic events do not fully capture the richness of real turbulent event data. These limitations place the benchmark in a specific methodological niche rather than making it a universal restoration dataset.

Availability is described cautiously. The paper states: "Demo code and data have been uploaded in supplementary material and will be released once accepted." At the time reflected in the manuscript text, the dataset is therefore intended for release but not yet publicly released.

A plausible implication is that the EGTM dataset’s importance lies less in scale alone than in modality design. It provides the paired sensing configuration necessary to test the claim that event density inversely correlates with turbulence-free lucky regions. In that sense, the dataset is both a benchmark and an experimental instrument for establishing whether event-guided turbulence mitigation is a viable alternative to frame-only TM.