Dark-EvGS: 3D Gaussian Splatting in Low Light

• Dark-EvGS is an event-assisted framework that reconstructs 3D radiance fields by fusing high-dynamic-range event data with long-exposure dark frames.
• It introduces triplet-level supervision—holistic, event, and mixed-sharpening losses—to enhance detail, reduce motion blur, and enforce color consistency.
• The method achieves superior performance with PSNR >20 and SSIM >0.8 in <40 lux conditions, while offering fast training and efficient GPU utilization.

Dark-EvGS is an event-assisted 3D Gaussian Splatting framework for low-light radiance-field reconstruction that enables the synthesis of bright frames from arbitrary viewpoints along the camera trajectory. It is formulated for regimes in which conventional cameras suffer from dynamic-range limitations and long-exposure motion blur, while event cameras provide high-dynamic-range, high-speed signals that remain essentially motion-blur-free even in near-dark conditions. The method combines raw event streams, synchronous long-exposure dark RGB frames, and known camera poses, introduces triplet-level supervision to couple holistic appearance, granular event structure, and sharpening constraints, and adds a Color Tone Matching Block to enforce color consistency. It is also accompanied by a real-captured dataset for event-guided bright-frame synthesis via 3D GS-based radiance-field reconstruction (Wu et al., 16 Jul 2025).

1. Problem formulation and methodological setting

Dark-EvGS addresses multi-view imaging under lighting below $40$ lux, where standard frame cameras exhibit two principal failure modes: limited dynamic range and severe motion blur induced by long exposure. The framework is motivated by the complementarity between event cameras and 3D Gaussian Splatting. Event cameras record per-pixel brightness changes asynchronously, with microsecond latency and high dynamic range $(120\ \mathrm{dB}+)$, whereas 3D Gaussian Splatting represents a scene as a set of 3D Gaussians with anisotropic covariance, spherical-harmonic color coefficients, and opacity, and renders novel views by splatting those Gaussians onto the image plane (Wu et al., 16 Jul 2025).

Within this setting, the event stream functions as the temporally precise and high-dynamic-range modality, while GS provides the radiance-field substrate that fuses multi-view information into a coherent 3D representation. The target output is not merely denoised or deblurred observations from the input viewpoints, but bright and sharp view synthesis at arbitrary viewpoints along the camera trajectory. The method is explicitly positioned as the first event-assisted 3D GS framework for this task, and its design responds to three low-light difficulties stated in the source formulation: noisy events, low-quality frames, and inconsistent color tone (Wu et al., 16 Jul 2025).

2. System architecture and radiance-field representation

The pipeline takes as input a raw event stream $E_{raw}$, synchronous long-exposure dark RGB frames $I_{low}$, and known poses $\{p_k\}$ for both frames and event timestamps. Each long-exposure frame $I_{low}^{t_k}$ is aligned with the accumulation of events over the interval $[t_k-w,t_k]$. Events are denoised with a $y$-noise filter, yielding $E_{gt}=Y(E_{raw})$. A frame encoder and an event encoder extract modality-specific features, which are fused by a Multimodal Coherence Modeling block that also predicts a global exposure parameter. The resulting decoded features are then processed by the Color Tone Consistency Module to produce pseudo bright frames $B^{t_k}$ (Wu et al., 16 Jul 2025).

For radiance-field reconstruction, Dark-EvGS initializes approximately $(120\ \mathrm{dB}+)$0 Gaussian primitives with learnable position, covariance, spherical-harmonic color coefficients, and opacity. The Gaussian kernel is defined as

$(120\ \mathrm{dB}+)$1

with covariance parameterized by

$(120\ \mathrm{dB}+)$2

Rendering at pose $(120\ \mathrm{dB}+)$3 proceeds by splatting Gaussians onto the image plane and blending them in depth order:

$(120\ \mathrm{dB}+)$4

Event supervision is integrated directly into GS. Given two rendered bright images $(120\ \mathrm{dB}+)$5 and $(120\ \mathrm{dB}+)$6, the predicted event representation is the log-intensity difference

$(120\ \mathrm{dB}+)$7

with $(120\ \mathrm{dB}+)$8 and $(120\ \mathrm{dB}+)$9. This $E_{raw}$0 is supervised against the filtered event accumulation $E_{raw}$1, which couples the 3D scene representation to the event modality rather than treating event processing as a separate pre-reconstruction stage (Wu et al., 16 Jul 2025).

3. Triplet-level supervision

A central component of Dark-EvGS is triplet-level supervision, introduced to address noisy events, imperfect pseudo bright frames, and motion blur through three complementary objectives. The first is the frame-level holistic loss,

$E_{raw}$2

which constrains the global shape and color layout through the pseudo bright frames. This provides scene-level appearance guidance from the fused event-frame representation (Wu et al., 16 Jul 2025).

The second objective is the event-level granular loss,

$E_{raw}$3

Its role is to sharpen fine edges and high-frequency detail that are not reliably recoverable from low-SNR dark frames alone. Because event measurements encode brightness changes rather than absolute radiance, this supervision is complementary to the holistic frame loss rather than redundant with it (Wu et al., 16 Jul 2025).

The third objective is the mixed-modality sharpening loss,

$E_{raw}$4

This term is motivated by the observation, stated in the source description, that subtraction of blurred frames yields sharper residuals. It therefore counteracts motion blur by comparing temporal differences between rendered images and pseudo bright frames in log space. The total loss is

$E_{raw}$5

The triplet design is thus organized around three scales of supervision: holistic appearance, granular event detail, and sharpening through cross-temporal differencing (Wu et al., 16 Jul 2025).

4. Color-tone consistency and pseudo-bright-frame generation

Dark-EvGS includes a Color Tone Matching Block within the Color Tone Consistency Module to address color inconsistency in rendered frames. The block is built upon channel-wise self-attention. Starting from the last decoded feature map $E_{raw}$6, a $E_{raw}$7 convolution and a depth-wise convolution generate query $E_{raw}$8, key $E_{raw}$9, and value $I_{low}$0. The tensors are reshaped so that $I_{low}$1 and $I_{low}$2 are mapped to $I_{low}$3, while $I_{low}$4 is mapped to $I_{low}$5, and an attention map is computed as

$I_{low}$6

The output feature is then reconstructed as $I_{low}$7, reassembled to $I_{low}$8, and combined with the predicted exposure from the Multimodal Coherence Modeling block to generate a globally consistent, locally refined pseudo bright frame $I_{low}$9 (Wu et al., 16 Jul 2025).

A notable design detail is that there is no additional color-consistency loss beyond $\{p_k\}$0; the CTMB is trained end-to-end within the triplet supervision. In this respect, color-tone control is handled architecturally rather than by introducing a dedicated auxiliary loss. The reported ablation results support the importance of this choice: removing CTMB on the “House” scene under dark $\{p_k\}$1 lux reduces PSNR from $\{p_k\}$2 to $\{p_k\}$3, SSIM from $\{p_k\}$4 to $\{p_k\}$5, and increases LPIPS from $\{p_k\}$6 to $\{p_k\}$7, with the source text explicitly noting color tone drift when CTMB is dropped (Wu et al., 16 Jul 2025).

5. Dataset, acquisition protocol, and optimization

The real-captured dataset introduced for Dark-EvGS uses a Davis 346 Color event camera fixed on a tripod, with a static object placed on a motorized turntable performing a $\{p_k\}$8 sweep. Each scene is captured under two lighting conditions. The bright condition, approximately $\{p_k\}$9 lux, provides sharp, high-SNR RGB frames and clean events. The dark condition, below $I_{low}^{t_k}$0 lux, produces long-exposure blurred RGB images and very noisy events; it includes three moderate scenes at $I_{low}^{t_k}$1–$I_{low}^{t_k}$2 lux and three challenging scenes at $I_{low}^{t_k}$3 lux. A TES 1339R light meter monitors illuminance, and hardware timestamps align each frame with an event interval $I_{low}^{t_k}$4. Camera intrinsics and hand-eye extrinsics are calibrated once per setup. The dataset comprises $I_{low}^{t_k}$5 scenes at approximately $I_{low}^{t_k}$6 fps over a $I_{low}^{t_k}$7 sweep, yielding approximately $I_{low}^{t_k}$8 frames and corresponding event streams per lighting condition (Wu et al., 16 Jul 2025).

The experimental setup uses an NVIDIA RTX 3090. Initialization employs $I_{low}^{t_k}$9 Gaussians uniformly random in the object’s bounding box. EvLowLight modules are pre-trained on $[t_k-w,t_k]$0 unrelated dark-bright pairs for $[t_k-w,t_k]$1 epochs and then frozen, with only CTMB randomly initialized. GS optimization runs for $[t_k-w,t_k]$2 iterations per scene, with learning rates specified as $[t_k-w,t_k]$3 for positions, $[t_k-w,t_k]$4 for features, $[t_k-w,t_k]$5 for opacity, $[t_k-w,t_k]$6 for scale, and $[t_k-w,t_k]$7 for the rotation quaternion. Training uses batch size $[t_k-w,t_k]$8 view at a time and Adam$[t_k-w,t_k]$9. The total training time per scene is reported as approximately $y$0 minutes on RTX 3090 (Wu et al., 16 Jul 2025).

6. Empirical performance, ablations, and limitations

Dark-EvGS is evaluated against seven baselines: 3D GS on dark frames only, two-stage E2VID + 3D GS, EventNeRF, Ev-GS, E2GS, SweepEvGS, and EvLowLight. Performance is measured with PSNR, SSIM, and LPIPS. The reported quantitative summary states that only Dark-EvGS consistently reaches PSNR $y$1 dB and SSIM $y$2, with LPIPS $y$3, across both moderate scenes (“Baseball,” “House,” “Lion”) and challenging scenes (“Panda,” “Badminton,” “Cat”). In efficiency terms, the method trains in approximately $y$4 minutes, renders at $y$5 FPS, and uses $y$6 GB of GPU memory on moderate scenes. By comparison, EventNeRF requires $y$7–$y$8 hours of training, renders at $y$9 FPS, and uses $E_{gt}=Y(E_{raw})$0 GB of memory, while other GS-based methods are reported at $E_{gt}=Y(E_{raw})$1–$E_{gt}=Y(E_{raw})$2 minutes of training, $E_{gt}=Y(E_{raw})$3–$E_{gt}=Y(E_{raw})$4 FPS, and $E_{gt}=Y(E_{raw})$5–$E_{gt}=Y(E_{raw})$6 GB of memory (Wu et al., 16 Jul 2025).

The ablation study on the “House” scene under dark $E_{gt}=Y(E_{raw})$7 lux attributes distinct functions to the major components. Full Dark-EvGS achieves PSNR $E_{gt}=Y(E_{raw})$8, SSIM $E_{gt}=Y(E_{raw})$9, and LPIPS $B^{t_k}$0. Removing CTMB yields $B^{t_k}$1, removing event supervision yields $B^{t_k}$2, removing the mixed-sharpening loss yields $B^{t_k}$3, and using only $B^{t_k}$4 yields $B^{t_k}$5. The accompanying interpretations are explicit: dropping CTMB causes a $B^{t_k}$6 dB PSNR loss and color tone drift; removing event loss degrades fine detail, with LPIPS increasing by $B^{t_k}$7; and ablating the mixed-sharpening loss produces blurrier renders (Wu et al., 16 Jul 2025).

The scope of the current method is also sharply delimited. Dark-EvGS is described as the first framework to fuse real dark events and dark frames into a 3D GS radiance field for synthesizing bright, sharp novel views in $B^{t_k}$8 lux, and the released dataset is identified as the first real-world low-light event-plus-frame multi-view dataset with $B^{t_k}$9 paired bright/dark scenes. At the same time, the stated limitation is that the present formulation supports only static scenes and does not handle deforming or highly non-rigid subjects. The indicated future direction is extension to dynamic radiance fields in the dark, which marks the principal boundary of the method as reported (Wu et al., 16 Jul 2025).