EventSeeColor: Learning to Reconstruct Color from Monochrome Events

Chuanzhi Xu; Langyi Chen; Haoxian Zhou; Zelin Zhang; Haodong Chen; Qiang Qu; Xiaoming Chen; Weidong Cai

cs.CVOct 29, 2026

EventSeeColor: Learning to Reconstruct Color from Monochrome Events

Authors: Chuanzhi Xu, Langyi Chen, Haoxian Zhou, Zelin Zhang, Haodong Chen, Qiang Qu, Xiaoming Chen, Weidong Cai

Organizations: The University of Sydney · Beijing Technology and Business University

Abstract

Event cameras capture visual brightness changes under challenging lighting conditions. However, traditional event-based image reconstruction targets grayscale intensity and produces images without scene color. Color distinguishes surfaces with similar intensity. Reconstructing color from events can reveal appearance differences that reconstructed grayscale images cannot express. Although monochrome event cameras do not directly measure color, their event-triggering patterns depend on surface reflectance, illumination, and the sensor's spectral response. These patterns may contain cues that support color reconstruction. In this paper, we present EventSeeColor, which uses event-reconstructed grayscale and aligned events to jointly guide a diffusion colorizer. Grayscale-derived semantic context supports color inference, while region-balanced chroma supervision and local RGB refinement improve color distribution, brightness, and texture. In the combined evaluation on two real event datasets, EventSeeColor achieves the best structural similarity, learned perceptual scores, and chromatic-region color difference among the external methods.

Figures & tables

Figure 2: EventSeeColor architecture. (a) Overall pipeline. (b) EGF fuses grayscale and event features. (c) BCR refines chroma and renders intermediate RGB. (d) LDR predicts a bounded RGB residual. (e) The diffusion colorizer uses ControlNet for spatial guidance and low-rank adaptation (LoRA) to adapt text cross-attention. EGF, BCR, and LDR apply validity masks before encoding.

Figure 3: Excluded RGB targets with their corresponding $50$ ms event accumulations. Each pair uses the same crop and target time. RGB values are displayed without exposure adjustment.

Image colorization from reconstructed grayscale
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$	Chr. $\Delta E_{00}\downarrow$	Desat. (%) $\downarrow$	CF-MAE $\downarrow$	NIQE $\downarrow$	FID $\downarrow$	Params. (M)
Colorful Image Colorization (ECCV, 2016) [ 66 ]	14.306	0.5185	0.5210	0.3073	22.876	11.114	19.054	6.766	207.114	32.24
DDColor (ICCV, 2023) [ 26 ]	13.705	0.4919	0.5224	0.3081	24.317	11.492	26.032	6.814	175.876	227.87
ColorFormer (ECCV, 2022) [ 25 ]	13.992	0.4862	0.4820	0.2975	21.798	18.978	18.210	6.850	138.680	44.82
BigColor (ECCV, 2022) [ 27 ]	14.380	0.4887	0.4865	0.2985	22.365	17.408	18.944	6.730	138.394	585.51
Control Color (IJCV, 2025) [ 34 ]	14.214	0.5962	0.4501	0.2881	20.841	22.304	20.150	7.222	156.302	1523.61

Table 1: Quantitative comparison with external baselines on PEOD and CoSEC. Chr. denotes chromatic target regions, and Desat. reports the percentage of chromatic target pixels reconstructed as near-neutral. CF-MAE measures colorfulness error relative to the RGB targets.

Figure 4: Qualitative comparison with the external baselines in Table 1 . EventSeeColor recovers more faithful surface colors and details while reducing false coloration on neutral regions across vegetation, buildings, roads, and vehicles.

Components and event guidance
Configuration	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$	Chr. $\Delta E_{00}\downarrow$	Desat. (%) $\downarrow$	CF-MAE $\downarrow$	NIQE $\downarrow$	FID $\downarrow$	Params. (M)
w/o EGF Event Path	15.981	0.6415	0.3371	0.2619	15.906	34.511	5.597	6.414	121.923	1523.91
w/o BCR	16.042	0.6439	0.3478	0.2638	16.398	45.884	6.383	6.416	100.750	1523.95
w/o LDR	15.056	0.6187	0.4113	0.3104	16.360	23.602	6.208	6.424	151.809	1523.96
w/o Event Inputs (EGF/BCR/LDR)	16.080	0.6357	0.3506	0.2617	16.229	40.126	5.618	6.333	131.985	1523.90
EventSeeColor (full)	16.010	0.6443	0.3329	0.2589	15.767	36.603	5.494	6.326	117.286	1523.98

Table 2: Ablation study of EventSeeColor’s components, event guidance, and training objectives on the evaluation set. Metrics and parameter-counting scope follow Table 1 . Within each section, best and second-best quality scores are bold and underlined, respectively.

Figure 5: EventSeeColor reconstructions of concrete, foliage, sky, and road surfaces. Rows show RGB targets, outputs (Ours), and corresponding close-ups. Fluorescent green boxes mark the enlarged regions. Crops and display scales match.

Figure 6: Target-color accuracy of EventSeeColor. Bars show crop-averaged regional CIEDE2000. Lower is better. Appendix C.3 defines the groups and aggregation.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Partition	Sequences	Frames	Crops
PEOD	Training	18	1,685	7,927
	Evaluation	5	479	2,198
CoSEC	Training	3	70	141
	Evaluation	1	12	22
Total	Training	21	1,755	8,068
	Evaluation	6	491	2,220

Appendix

Table A1 : Data used for training and evaluation. Sequences are disjoint between the two partitions within each dataset.

Component	Setting	Value
Inputs	Aligned crop / diffusion image	$256\times 256$ / $512\times 512$
	Event bins per polarity / total channels	$10$ / $24$
	Voxel scale / count scale	$\tau_{\rm voxel}=16$ , $\tau_{\rm count}=64$
EGF	Stage widths / strides	$(16,32,64)$ / $(1,2,2)$
	Stage operator / normalization	$3\times 3$ convolution, 4-group GroupNorm, SiLU
	Gate and projection kernels / hint size	$1\times 1$ / $320\times 64\times 64$

Appendix

Table A2 : Network and inference settings. Lightness calibration is specified in Appendix B.1 .

Component	Setting	Value
Chroma objectives	Chroma scale / pixel weighting	$s=32$ , $\eta=2$ , $\kappa=20$
	Region thresholds / warm hue	$\tau_{n}=5$ , $\tau_{c}=10$ , $[h_{-},h_{+}]=[-30^{\circ},75^{\circ}]$
	Auxiliary chroma / region weights	$\lambda_{\rm aux}=0.1$ , $\beta=0.5$
	Shortfall / neutral / warm weights	$\lambda_{s}=0.2$ , $\lambda_{n}=0.5$ , $\lambda_{w}=0.2$
LDR objective	RGB / gradient / color weights	$\lambda_{\rm rgb}=1$ , $\lambda_{\rm grad}=0.1$ , $\lambda_{\rm color}=0.25$
	Perceptual / clipping weights	$\lambda_{\rm perc}=0.1$ , $\lambda_{\rm clip}=0.1$

Appendix

Table A3 : Training settings and objective weights. Region thresholds are expressed in Lab chroma units.

Figure A1 : Additional comparisons with external baselines. Rows (a) - (c) show vegetation, (d) - (f) road markings and texture, and (g) - (i) a facade, guardrail, and vehicle. Native RGB outputs use identical crop coordinates and display scales, with the RGB target as reference.

Method	FID $\downarrow$	KID $\times 10^{3}$ $\downarrow$	CF-MAE $\downarrow$	All $\Delta E_{00}$ $\downarrow$	Neutral $\Delta E_{00}$ $\downarrow$	NIQE $\downarrow$	BRISQUE $\downarrow$
Image colorization from reconstructed grayscale
Colorful Image Colorization	207.114	151.511 $\pm$ 2.959	19.054	22.490	21.671	6.766	14.711
DDColor	175.876	110.633 $\pm$ 2.765	26.032	23.284	22.345	6.814	14.688
ColorFormer	138.680	78.946 $\pm$ 2.359	18.210	20.664	20.126	6.850	14.378
BigColor	138.394	80.011 $\pm$ 2.439	18.944	21.000	20.380	6.730	14.469
Control Color	156.302	94.606 $\pm$ 2.868	20.150	20.301	19.762	7.222	27.166

Appendix

Table A4 : Distribution, full-image color difference, and auxiliary appearance metrics on native RGB outputs. KID mean and subset standard deviation are multiplied by

10^{3}

. Lower values are better for all metrics. The same input groups as Table 1 in the main paper apply.

Method	All $ab$ $\downarrow$	Chr. $ab$ $\downarrow$	Neutral $ab$ $\downarrow$	Warm $ab$ $\downarrow$	Desat. (%) $\downarrow$	False color (%) $\downarrow$	Warm leak (%) $\downarrow$
Image colorization from reconstructed grayscale
Colorful Image Colorization	14.839	14.252	13.963	12.632	11.114	34.237	6.347
DDColor	16.170	16.626	15.267	14.954	11.492	38.603	21.012
ColorFormer	11.870	13.402	11.221	14.960	18.978	18.761	8.791
BigColor	14.921	15.956	14.050	16.398	17.408	33.901	1.575
Control Color	12.662	13.687	11.824	12.962	22.304	27.229	17.823

Appendix

Table A5 : Chroma error and regional color behavior of native RGB outputs. Chroma RMSE occupies the first four numeric columns. The final columns report desaturation, neutral false color, and warm color leakage in percent. Lower is better.

Panel of Figure 14 — Figure A2 : Qualitative component and event-guidance comparisons. The configurations correspond to Table 2 in the main paper. Rows (a) and (b) show foliage, row (c) shows a guardrail, and row (d) shows a roadside vehicle. Each row uses the same crop and scale.

Configuration	$ab$ RMSE $\downarrow$	Approx. $\Delta E_{76}\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$
Frozen Control Color	12.8797	26.0718	14.3973	0.6007
LoRA only	9.7619	23.5487	14.6786	0.6193
LoRA + gray adapter	9.4248	23.1856	14.7649	0.6194
LoRA + event–gray fusion	9.1218	22.9109	14.8167	0.6209
EventSeeColor (full)	4.7769	19.5217	15.1163	0.6176

Appendix

Table A6 : Data adaptation, fusion, and the complete method with common fixed-lightness outputs on 2,220 evaluation crops. Best and second-best values are bold and underlined.

Event-Based Vision

Event-based vision uses asynchronous camera measurements to perceive motion and scene changes with high temporal resolution and dynamic range.

See all 236 papers

SEE Challenge 2026: Event-Guided Brightness Adjustment Across a Broad Illumination Range
Sep 24, 2026Yunfan Lu, Mingchao Xu, Hanyu Zhou +9Image Enhancement HDR Imaging
Hong Kong University of Science and Technology (Guangzhou) · National University of Singapore · Xidian University +5
4D Reconstruction with Events & RGB
Oct 27, 2026Dulanga Weerakoon, Darshana Rathnayake, Van Minh Do +74D Reconstruction RGB-Event Sensor Fusion
Singapore-MIT Alliance for Research & Technology Centre · Singapore Management University · Massachusetts Institute of Technology
Event-based Continuous Color Video Decompression from Single Frames
Nov 30, 2023Ziyun Wang, Friedhelm Hamann, Kenneth Chaney +3Event-Based Vision
University of Pennsylvania, USA. · TU Berlin, ECDF Berlin, SCIoI Cluster and RIG, Germany. · Archimedes, Athena RC

Sep 24, 2026J/K move · Enter open · S save

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

Yunfan Lu, Mingchao Xu, Hanyu Zhou, Shaoyu Liu+8

Hong Kong University of Science and Technology (Guangzhou) · National University of Singapore · Xidian University+5

Image Enhancement HDR Imaging Low-Light Image Enhancement Image Restoration

Open paper AskPDF arXiv

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.

Abstract

Event-Based Vision

Explore similar work

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

4D Reconstruction with Events & RGB

Event-based Continuous Color Video Decompression from Single Frames