PaperlayerView exact revision on arXiv

Source-linked AI summary

SEE Challenge 2026: Event-Guided Brightness Adjustment Across a Broad Illumination Range

Yunfan Lu, Mingchao Xu, Hanyu Zhou, Shaoyu Liu, Haoyue Liu, Peiqi Duan, Shihan Peng, Yinqiang Zheng, Boxin Shi, Gim Hee Lee, Hui Xiong, Davide Scaramuzza

arXiv:2609.29347v1cs.CV

TL;DR

Conventional frames may lose useful measurements under broad illumination, and brightness adjustment can map one input to different valid references unless its target brightness is specified. The SEE Challenge 2026 evaluates verified event-guided restoration systems under an open-system protocol; top aggregate scores are closely spaced, while results vary by case and metric.

  • Problem

    Under broad illumination, RGB frames may be degraded, and the same input can correspond to different valid reference images unless target brightness is specified.

    Source: §1 · p. 1, §2.4 · p. 5

  • Method

    The SEE Challenge 2026 evaluates organizer-verified event-guided restoration systems using RGB-event inputs conditioned on a target-brightness statistic under an open-system protocol.

    Source: §1 · p. 3, §2.4 · p. 5, §2.5 · p. 6

  • Results

    The top two systems differ by only 0.0858 dB PSNR, while the best method varies across cases and metrics and severe underexposure leaves visible local errors.

    Source: §3 · pp. 7–8, §6 · p. 14

  • Takeaways & Limitations

    The reported ranking compares complete submitted pipelines and does not isolate the contribution of events or any individual system component.

    Source: §1 · p. 3

    Because systems differ in frame counts, future context, pretraining, test-time augmentation, and post-processing, the ranking cannot isolate individual component contributions.

    Source: §5 · p. 13

Abstract

from arXiv · show

Event cameras provide a high dynamic range and preserve brightness-change cues in lighting conditions where conventional RGB frames may be noisy or saturated. To benchmark event-guided restoration across a broad illumination range, we organized the SEE Challenge 2026 with the Event-Based Multimodal Vision Workshop at ECCV 2026. The task conditions restoration on one or more RGB frames, synchronized events, and a scalar target-brightness statistic provided by the organizers. It uses SEE-600K, which contains 610,126 image-event observations from 202 real-world scenes spanning low-light, normal-light, and high-light conditions with illumination variations of up to 1,000$\times$. The challenge follows an open-system protocol: participants may use different temporal contexts, architectures, pretrained weights, test-time augmentation, and post-processing strategies. PSNR determines the ranking, and SSIM is reported as a secondary metric. Around 70 teams registered interest and 15 valid CodaBench submissions were received. Six distinct teams completed organizer-side identity and technical verification, provided method descriptions, checkpoints, inference code, and instructions, and are included in the verified open-system ranking reported here. Beyond the ranking, this report analyzes exposure subsets, semantically distinct test cases, a shared failure pattern, system design choices, and inference strategies. The top systems obtain closely spaced average scores, while the best-performing method varies across cases and metrics; under severe underexposure, all verified systems retain visible local errors.

1 Introduction

Conventional frames can lose useful structure and color under extreme illumination, while event cues offer complementary brightness-change information. The SEE Challenge 2026 benchmarks diverse event-guided restoration systems and analyzes their results and design choices.

1 Introduction

  • SEE-600K spans low-, normal-, and high-light conditions, whereas SDE covers low-to-normal brightness; the challenge conditions restoration on reference-derived target brightness B.Earlier low-light methods map dark inputs to a fixed normal-light target [4]; SEE-600K is described as spanning the broader range [15].Source: Figure 1 · p. 2
  • Under degraded illumination, dark regions can be noise-dominated and bright regions saturated, making faithful structure and color recovery from RGB alone difficult.Event cameras preserve brightness-change cues but do not directly provide dense color or absolute brightness, motivating fusion with RGB and a specified target exposure.Source: §1 · p. 2
  • Around 70 teams registered interest, 15 valid submissions were received, and six distinct teams completed technical verification for the reported study.The systems span cross-attention, feature modulation, state-space models, latent flow, temporal averaging, cascaded restoration, and test-time augmentation.Source: §1 · p. 3
  • The challenge report defines the task and protocol, ranks six verified systems, analyzes exposure and content cases, and compares system and inference choices.Because systems vary in temporal context, pretrained weights, test-time augmentation, cascades, and post-processing, the ranking evaluates complete pipelines rather than individual components.Source: §1 · p. 3

2 Challenge Protocol

The challenge defines RGB-event restoration toward a specified target brightness and evaluates it under an open-system protocol. Its protocol also specifies how data, event intervals, and organizer-side verification support the benchmark.

2.1 Task Definition and Open-System Scope

  • Systems map one or more RGB frames, associated events, and organizer-provided target-brightness statistic B to a restored image for Low-to-Normal or High-to-Normal restoration.The open-system rules impose no maximum temporal context and permit participants to choose their event representation, fusion strategy, training objective, and inference pipeline.Source: §2.1 · pp. 3–4

2.2 Dataset Splits and Pair Construction

  • SEE-600K contains 610,126 image-event observations from 202 scenes, while the released splits yield 656,500 constructed pairs and the final evaluation uses six newly recorded scene groups.The six groups' normal-light references remain hidden; constructed restoration pairs can outnumber distinct observations because multiple source-target mappings may be formed.Source: §2.2 · p. 4

2.3 Event–Frame Synchronization and Data Format

  • Each frame’s events are grouped from its exposure-start timestamp to the next frame’s, using a half-open interval; released inputs include 8-bit RGB PNGs and timestamped event arrays.Repeated trajectories are registered using synchronized 1-kHz IMU signals; the dataset paper reports 0.2967-pixel mean displacement and at most approximately 1 ms temporal registration error.Source: §2.3 · pp. 4–5

2.4 Target-Brightness Statistic

  • B is the per-target-frame arithmetic mean of normalized sRGB values across pixels and RGB channels, computed by organizers from the hidden reference and supplied at test time.It is ground-truth-derived information, and results do not evaluate generalization to arbitrary user-selected brightness values.Source: §2.4 · p. 5
  • Because one RGB-event input can correspond to multiple valid target brightnesses, conditioning on B selects the desired brightness and reduces target ambiguity under the stated squared reconstruction loss.Without B, the optimal squared-loss prediction is the conditional mean reference given the input; conditioning makes it the mean given both input and B.Source: §2.4 · pp. 5–6

2.5 Evaluation Protocol and Challenge Rules

  • PSNR determines the verified ranking, SSIM is secondary, and the reported average is the mean of the Low-to-Normal and High-to-Normal subset scores.Phase 2 CodaBench samples one frame in ten, while organizer-side technical verification uses dense temporal sampling.Source: §2.5 · p. 6, Table 2 · p. 7
  • Fifteen valid CodaBench submissions were received, but only six distinct teams completed identity and technical verification and provided materials for the organizer-side ranking.A platform account alone does not establish a distinct verified team identity.Source: §2.5 · p. 6

2.6 Prior Evidence and Official Baseline

  • The released SEE-Net baseline uses one RGB frame, a synchronized 64-bin event voxel grid, and a brightness statistic, scoring 18.8279 dB PSNR and 0.6414 SSIM on the released test split.These released-split results are provided for reference.Source: §2.6 · p. 7

3 Challenge Results and Organizer-Side Analysis

The verified systems achieve closely spaced average scores, but their relative strengths vary across exposure settings, content, and metrics. Severe underexposure remains a shared challenge: systems brighten the input without faithfully recovering checkerboard detail.

Verified Open-System Ranking

  • MLSLabs ranks first under dense sampling at 23.1731 dB PSNR and 0.7237 SSIM, followed by ACVLab-TL at 23.0873 dB PSNR and 0.7169 SSIM.The 0.0858 dB PSNR gap is too small to support a claim of clear separation.Source: §3 · pp. 7, 9
  • Dense sampling changes subsampled scores by less than 0.04 dB PSNR and 0.001 SSIM across all six systems, closely estimating dense-sequence averages.Organizer-side runs also verify that the supplied code and checkpoints reproduce submitted predictions.Source: §3 · pp. 7–8

Representative Case Analysis

  • Robot Motion (G1) has the lowest PSNR for every system; ACVLab-TL leads its metrics at 18.85 dB PSNR and 0.5010 SSIM.One group cannot isolate motion from illumination, scene content, event density, or registration effects.Source: Table 3 · p. 8, §3 · p. 8

Performance Across Exposure Settings

  • For pixelartai, Low-to-Normal PSNR is 21.1863 dB versus 25.1600 dB for High-to-Normal, making the former harder in this test composition.The score gap is consistent with weaker, noisier underexposed RGB, but aggregate scores do not establish that mechanism.Source: §3 · p. 8

Representative Case Analysis

  • Across G2 and G3, metric leaders differ; Chinese Text yields each system’s highest case PSNR, and all submissions outperform SEE-Net in all three groups.Checkerboard texture and Chinese text produce different rankings across pixel-fidelity, structural, and perceptual metrics.Source: §3 · p. 8, Table 3 · p. 8

Organizer-Side Synthesis

  • The open-system ranking cannot attribute outcomes to individual design choices, while disclosed configurations support hypotheses about temporal context, TTA, and model scale.Event-SCAM alone aggregates adjacent-frame outputs; the other five systems use a single target-frame input, even when they repeat inference.Source: §3 · p. 9, Table 4 · p. 12

Qualitative Observations

  • Fig. 2 shows residual noise, color shifts, weakened texture, and local-contrast errors despite greater visible structure than the released baseline.The examples use a common image range without extra gamma or contrast adjustment; their differences are not fully captured by average PSNR and SSIM.Source: §3 · p. 9, Figure 2 · p. 10

Shared Failure Case

  • Under severe underexposure, no system faithfully reconstructs checkerboard structure; best PSNR is 18.73 dB, best SSIM 0.5200, and best LPIPS 0.2688.The best values belong to different systems; the methods largely correct global brightness but retain blurred boundaries, weakened contrast, and color shifts.Source: Figure 3 · p. 11, §3 · p. 11

4 Participating Methods

The verified submissions span different RGB–event fusion, brightness-conditioning, temporal-processing, and inference strategies. Only Event-SCAM aggregates adjacent-frame outputs; the other five systems use a single target frame.

4 Participating Methods

  • Table 4 compares pretraining, brightness conditioning, test-time augmentation, temporal processing, and inference multiplicity across the six verified systems.It distinguishes multi-frame temporal processing from repeated inference on one target frame.Source: §4 · p. 12
  • Event-SCAM uses bidirectional cross-attention and conditionally averages adjacent-frame outputs, while EventRestormer combines spatial FiLM fusion with a two-stage cascade and eight-way same-frame TTA.Event-SCAM uses up to seven single-frame calls; EventRestormer makes 16 stage-level passes without adjacent frames.Source: §4 · pp. 12–13
  • EventFuse combines Bayer and coordinate encodings, a Mamba-based illumination estimator, bidirectional fusion, and sparse state-space blocks in a 1.16M-parameter model.It uses one RGB frame and a 64-bin event grid, with no external data or pretrained weights.Source: §4 · p. 12
  • EvLCD predicts an exposure-corrected base and event-guided residual, then averages four flipped same-frame predictions and shifts their global mean to match B.It is trained from scratch on official data in two stages.Source: §4 · p. 13
  • REFM performs six Euler updates in a learned RGB latent space, while SEE-SplitNet selects an exposure-specific model and applies four same-frame TTA variants.REFM uses an exposure descriptor and binary event voxel grid; SEE-SplitNet trains separate Low-to-Normal and High-to-Normal variants.Source: §4 · p. 13

5 Limitations

The verified ranking compares complete open-system pipelines rather than isolating individual components. Its evaluation also has limits in target-brightness control, temporal and statistical analysis, and cost comparability.

5 Limitations

  • Because systems differ in frame counts, future context, pretraining, cascades, TTA, temporal averaging, and post-processing, the ranking cannot isolate the contribution of events or any individual component.No matched RGB-only retraining was conducted on the hidden set, so a fair organizer-run modality ablation was unavailable.Source: §5 · pp. 13–14
  • The target-brightness statistic is derived from each hidden reference, unlike practical user-selected brightness, and the challenge lacks a standardized quantitative sweep across methods.The missing sweep concerns monotonicity and output-brightness error.Source: §5 · p. 14
  • Frame-wise PSNR and SSIM, selected representative cases, and one shared failure example do not provide a complete per-sequence distribution or temporal-consistency evaluation.The challenge also lacks systematic analysis by event rate or scene motion and scene-level bootstrap confidence intervals; small score differences are not statistically established.Source: §5 · p. 14
  • Reported computation and latency are not controlled end-to-end comparisons because teams used different hardware and software, may omit pipeline stages, and incur input-dependent temporal-averaging costs.The reported FLOPs describe core networks rather than standardized end-to-end cost.Source: §5 · p. 14

6 Conclusion

The SEE Challenge 2026 reports a verified open-system ranking for event-guided restoration conditioned on a target-brightness statistic. Its analyses find closely spaced average scores, case- and metric-dependent best methods, and persistent local errors under severe underexposure.

6 Conclusion

  • Six verified teams are included in the reported ranking, following 15 valid CodaBench submissions.The challenge is an open-system benchmark for target-brightness-statistic-conditioned restoration across a broad illumination range.Source: §6 · p. 14
  • Low-to-Normal restoration obtains lower scores, and the best-performing method varies across cases and metrics.Dense sampling produced results close to the subsampled server evaluation.Source: §6 · p. 14
  • Under severe underexposure, correcting global brightness does not ensure faithful recovery of fine local texture.The challenge identifies open questions in exposure recovery, efficiency, temporal consistency, and practical brightness control.Source: §6 · p. 14
Loading 2609.29347v1…