---
title: 'EeveeDark: Binary Low-Light Video Enhancement'
url: https://www.emergentmind.com/topics/eeveedark
type: topic
---

# EeveeDark: Binary Low-Light Video Enhancement

EeveeDark is a low-light video enhancement framework that combines sensor-level Bayer RAW frames with event streams inside a largely binary neural network, targeting the regime in which extreme darkness, temporal coherence, and resource-constrained deployment must be addressed simultaneously [2607.06217]. It is formulated for sequences of noisy, dark RAW video frames and seeks to recover clean, bright, temporally coherent outputs while retaining computational efficiency suitable for embedded or robotic platforms. The method is explicitly positioned around three coupled design choices: sensor-level fusion rather than processed RGB input, event-guided multimodal enhancement rather than image-only restoration, and binary neural computation rather than full-precision video restoration throughout most of the network [2607.06217].

## 1. Problem formulation and design rationale

EeveeDark targets **low-light video enhancement** in settings where extreme darkness causes low signal-to-noise ratio, low contrast, color degradation, and unstable temporal behavior [2607.06217]. The problem is not treated as a framewise enhancement task alone. Because the input is video, the method also seeks to avoid flicker and exploit temporal redundancy without introducing motion artifacts.

The framework is motivated by two shortcomings that the paper attributes to existing event-guided low-light enhancement methods. First, they often use **processed RGB inputs**, which lose sensor-level information important for faithful restoration in severe darkness. Second, they often rely on expensive alignment or attention modules, such as optical flow, which the paper identifies as unsuitable for edge deployment [2607.06217]. EeveeDark therefore adopts **sensor-level fusion**, starting from Bayer RAW and combining it with event data rather than enhancing already processed RGB imagery.

The two sensing modalities are treated as complementary. **RAW frames** provide spatial richness, photometric fidelity, and sensor-level color and tonal information. **Event streams** provide high temporal resolution, large dynamic range, and motion and brightness-change cues that remain informative in difficult lighting [2607.06217]. The paper’s central claim is that this multimodal input can compensate for the quality degradation ordinarily induced by binarization, thereby enabling a lightweight model that remains practical on resource-constrained hardware.

The binary design is not presented as an isolated efficiency device. Rather, the method argues that a **Binary Neural Network (BNN)** can become viable for restoration if the input representation itself is sufficiently informative. This suggests that EeveeDark should be understood less as a conventional BNN adapted to low light than as a sensor-fusion architecture in which binary computation is made workable by RAW-domain processing and event guidance.

## 2. Input representation and overall pipeline

At a high level, EeveeDark takes a sequence of **packed Bayer RAW frames** and synchronized **event streams** converted to voxel grids between frames, and predicts **enhanced RAW frames**, which are then converted to RGB using a standard ISP for visualization and evaluation [2607.06217].

For the RAW input, the paper defines a Bayer sequence
$$
\{\mathbf{I}_B^1,\mathbf{I}_B^2,\dots,\mathbf{I}_B^T\}, \qquad \mathbf{I}_B^i \in \mathbb{R}^{H\times W}.
$$
Each \(2\times2\) Bayer block is packed into four channels, yielding
$$
\mathbf{I}_p^i \in \mathbb{R}^{\frac{H}{2}\times \frac{W}{2}\times 4},
$$
and amplified using a scaling factor \(r\), although the exact value of \(r\) is not stated in the paper excerpt [2607.06217].

For events, each event is represented as
$$
e_i=(x_i,y_i,t_i,p_i),
$$
with polarity \(p_i\in\{+1,-1\}\). Events are grouped between consecutive frame times into sets \(G_k\), then converted into voxel grids
$$
E_k \in \mathbb{R}^{\frac{H}{2}\times \frac{W}{2}\times B}
$$
using the polarity-based linear accumulation method of Rebecq et al., with
$$
B=5
$$
temporal bins in all experiments [2607.06217].

The processing pipeline contains seven stages: preprocessing; modality-specific binary encoders; lightweight multimodal fusion; a shift encoder with recurrent embeddings; an Event-Guided Skip Gate (EGSG); a shift decoder with recurrent embeddings; and a residual restoration module that outputs enhanced RAW [2607.06217]. The architecture is described as having five key components: preprocessing and separate encoding of Bayer RAW and events; efficient multimodal fusion; a shift encoder carrying temporal information and using EGSG; a shift decoder refining and aligning temporal features; and a final restoration module whose output is later passed through an ISP.

This structure is explicitly distinct from a flow-based alignment pipeline. Temporal propagation is handled by grouped spatial-temporal shift operations from ShiftNet rather than optical-flow estimation [2607.06217]. A plausible implication is that the method’s efficiency claim depends not only on binarization, but also on replacing motion estimation with shift-based temporal exchange.

## 3. Network architecture

EeveeDark uses two separate encoders, one for frames and one for events, and the paper states that they employ **distribution-aware binary convolutions** from BRVE [2607.06217]. The first convolution in each encoder is kept **full precision** to preserve low-level detail from the RAW and event inputs. The encoder outputs are written as
$$
\mathbf{F}_t^{\text{frame}} = \text{BinaryEncoder}(\mathbf{I}_t), \qquad
\mathbf{F}_t^{\text{event}} = \text{BinaryEncoder}(\mathbf{E}_t).
$$

The fusion block combines the two modalities by channel concatenation:
$$
\mathbf{F}_t^{\text{concat}} = \text{Concat}\left(\mathbf{F}_t^{\text{frame}}, \mathbf{F}_t^{\text{event}}\right).
$$
The fused representation is then computed through a binary \(3\times3\) branch and a full-precision \(1\times1\) branch:
$$
\mathbf{F}_t^{\text{fused}} = \text{RPReLU}\Big(
\phi_{\text{binary}}(\mathbf{F}_t^{\text{concat}})
+
\phi_{\text{full}}(\mathbf{F}_t^{\text{concat}})
\Big).
$$
The paper characterizes this as a deliberately lightweight fusion mechanism rather than an attention-based cross-modal transformer [2607.06217].

Temporal propagation begins in a shift encoder with recurrent embeddings:
$$
\mathbf{T}_t^{\text{shifted}}=
\text{ChannelShift}\Big(
\text{Concat}(\mathbf{H}_{t-2},\mathbf{H}_{t-1},\mathbf{F}_t^{\text{fused}})
\Big),
$$
followed by
$$
\mathbf{F}_t^{\text{shifted}}=
\text{Downsample}\big(
\phi_{\text{binary}}(\mathbf{T}_t^{\text{shifted}})
\big).
$$
This design propagates temporal information using cyclic channel shift rather than explicit motion compensation [2607.06217].

A principal novelty is the **Event-Guided Skip Gate (EGSG)**, which uses event features to generate a gating map for the shifted encoder features. The event feature map is resized and projected:
$$
\mathbf{F}_t^{\text{proj}}=
\phi_{\text{proj}}\big(
\text{BilinearInterpolate}(\mathbf{F}_t^{\text{event}})
\big),
$$
then summarized through channel-wise statistics:
$$
\mathbf{Z}_t=
\text{Concat}\Big(
\text{mean}(\mathbf{F}_t^{\text{proj}}),
\text{mean}(|\mathbf{F}_t^{\text{proj}}|),
\text{std}(\mathbf{F}_t^{\text{proj}})
\Big).
$$
The gate is obtained as
$$
\mathbf{G}_t=
\sigma\big(
\phi_{\text{attention}}(\mathbf{Z}_t)
\big),
$$
and applied multiplicatively:
$$
\mathbf{F}_t^{\text{gated}}=
\mathbf{G}_t \odot \mathbf{F}_t^{\text{shifted}}.
$$
The interpretation given in the paper is that event-derived statistics indicate where brightness changes and motion occur, allowing the gate to emphasize dynamic regions and suppress redundant ones [2607.06217].

The decoder operates on a local temporal window \(\{t-2,t-1,t\}\). Each gated feature map is split into stationary and shiftable components:
$$
\left(\mathbf{F}_i^{(1)}, \mathbf{F}_i^{(2)}\right)=
\text{Split}(\mathbf{F}_i^{\text{gated}}),
\qquad i\in\{t-2,t-1,t\}.
$$
The shiftable parts are cyclically redistributed as
$$
\mathbf{S}_t=
\left\{
\mathbf{F}_{t}^{(2)},
\mathbf{F}_{t-2}^{(2)},
\mathbf{F}_{t-1}^{(2)}
\right\},
$$
and a spatial shift is applied using
$$
\mathbf{S}_t^{\text{spatial}}=
\text{SpatialShift}(\mathbf{S}_t,\mathcal{K}),
$$
with
$$
\mathcal{K}= \{(x,y)\mid x,y\in\{-8,-4,0,4,8\},\ (x,y)\neq(0,0)\}.
$$
The decoded feature is then
$$
\mathbf{F}_t^{\text{decoded}}=
\text{UpSample}\Big(
\phi_{\text{binary}}\big(
\text{Concat}(\mathbf{S}_t^{\text{spatial}},\mathbf{F}_t^{(1)})
\big)
\Big),
$$
and the final restoration is residual:
$$
\mathbf{O}= \mathbf{I}_p + \phi_{\text{out}}(\mathbf{F}_t^{\text{decoded}}).
$$
The output layer \(\phi_{\text{out}}\) is full precision [2607.06217].

The paper does not provide explicit quantization equations, straight-through estimator formulas, exact channel widths, or layer counts. It states only that most internal convolutions use binary weights and activations, while the first layer in each encoder, selected \(1\times1\) fusion and gating layers, and the final output layer remain full precision [2607.06217].

## 4. Training objective and implementation protocol

Training is performed entirely in the RAW domain using only the **Charbonnier loss**:
$$
\mathcal{L}_{\text{Charbonnier}}=
\sqrt{\|\hat{\mathbf{I}}-\mathbf{I}_{\text{GT}}\|^2+\epsilon^2},
\qquad \epsilon=10^{-3}.
$$
Here \(\hat{\mathbf{I}}\) denotes the predicted frame and \(\mathbf{I}_{\text{GT}}\) the ground-truth RAW frame [2607.06217]. The paper explicitly does not include perceptual loss, SSIM loss, temporal consistency loss, adversarial loss, or an event-specific auxiliary loss.

Implementation details are reported as follows: **PyTorch** is used for training on an **NVIDIA RTX A5000**; the training unit is **\(128\times128\) patches** from **10-frame sequences**; augmentations are **random cropping, horizontal flipping, vertical flipping**; optimization uses **Adam** with
$$
\beta_1=0.9,\quad \beta_2=0.99;
$$
training runs for **100,000 iterations**; and the learning rate follows **cosine annealing restart**, decaying from
$$
2\times10^{-4} \to 1\times10^{-7}.
$$
The batch size is not stated in the excerpt [2607.06217].

The efficiency analysis is expressed through explicit complexity formulas:
$$
\text{FLOPs}=\frac{\text{OPs}^{\text{bin}}}{64}+\text{OPs}^{\text{fp}},
$$
and
$$
\text{Params}=\frac{\text{Params}^{\text{bin}}}{32}+\text{Params}^{\text{fp}}.
$$
Average per-frame FLOPs are reported on a 100-frame video at \(256\times256\), while latency is estimated using **daBNN** on ARM64 mapping and energy through a **7nm device model** following prior BNN work [2607.06217]. The paper notes that direct GPU runtime is not meaningful because current GPU frameworks do not natively support binarized layers.

A methodological point follows from these choices. Because the loss remains purely photometric in the RAW domain despite the use of events, the event stream functions as an internal guidance modality rather than a source of explicit supervisory signal. This distinguishes EeveeDark from event-guided radiance-field methods such as Dark-EvGS, where event supervision appears directly in the optimization objective [2507.11931].

## 5. Datasets, evaluation settings, and empirical results

The experiments cover **LLRVD**, **HUE**, **SDE**, **SDSD**, and **CEAR** [2607.06217]. The paper states that methods on **LLRVD** and **HUE** were retrained from scratch using public implementations. **HUE** has no reference ground truth for enhancement, so no-reference IQA metrics are used there. **CEAR** frames are warped into the event camera frame using calibrated intrinsics and extrinsics for downstream evaluation [2607.06217].

The comparison settings are divided into three groups. For **RAW-to-RAW on LLRVD**, baselines are **BBCU**, **BRVE**, **ShiftNet**, and **FloRNN**. For **RGB-domain evaluation via ISP on LLRVD/HUE**, the comparisons include those methods plus **EvLight**. For **RGB benchmarks on SDE/SDSD**, the paper additionally compares against **SNR-Net**, **Uformer**, **Retinexformer**, **ELIE**, **eSL-Net**, **Liu et al.**, **EvLight**, **E2VID+**, and **BRVE** [2607.06217].

On **LLRVD** in the RAW domain, Table 1 reports:

| Method | PSNR | SSIM | ST-RRED |
|---|---:|---:|---:|
| FloRNN | 37.20 | 0.960 | 0.042 |
| ShiftNet | 37.85 | 0.967 | 0.031 |
| BBCU | 36.52 | 0.950 | 0.058 |
| BRVE | 37.07 | 0.958 | 0.046 |
| EeveeDark | **37.51** | **0.962** | **0.039** |

These numbers show that EeveeDark improves over the prior binary methods **BBCU** and **BRVE**, with a reported gain of **+0.99 dB PSNR** over BBCU and **+0.44 dB PSNR** over BRVE, while approaching the quality of the full-precision models [2607.06217]. The paper also reports that BRVE exhibits temporal flickering and that EeveeDark improves temporal stability.

After ISP on **LLRVD**, the paper reports **30.37 / 0.850 / 0.189** for EeveeDark, compared with **29.58 / 0.821 / 0.233** for BRVE, **28.72 / 0.795 / 0.373** for BBCU, **30.28 / 0.845 / 0.185** for EvLight, **30.53 / 0.853 / 0.176** for FloRNN, and **32.01 / 0.889 / 0.095** for ShiftNet [2607.06217]. On **HUE**, EeveeDark yields **0.177** CLIP-IQA, **0.158** MANIQA, **0.605** TOPIQ-NR, and **5.358** NIQE, with NIQE being the best among the listed methods [2607.06217].

On **SDE**, EeveeDark reports **21.99** PSNR, **24.32** PSNR\*, and **0.707** SSIM, compared with **21.52 / 23.46 / 0.672** for BRVE and **22.83 / 25.21 / 0.760** for EvLight [2607.06217]. On **SDSD**, it reports **26.66 / 28.69 / 0.859**, compared with **24.14 / 27.72 / 0.816** for BRVE and **27.60 / 30.02 / 0.875** for EvLight [2607.06217]. The paper’s qualitative interpretation is that BRVE oversmooths and amplifies noise in dark regions, whereas EeveeDark restores finer structures and local contrast using event information.

The ablation results isolate two central components. The **full EeveeDark (RAW + Events)** model obtains **37.51 / 0.962 / 0.039** on LLRVD; **w/o Event Encoder (RAW only)** drops to **37.07 / 0.958 / 0.046**; and **w/o EGSG** yields **37.44 / 0.962 / 0.041** [2607.06217]. The paper also compares enhancement settings: **RAW2RAW + ISP** gives **30.37 / 0.8501 / 0.2021**, **RAW2RGB** gives **28.26 / 0.8492 / 0.2393**, and **RGB2RGB** gives **26.79 / 0.8165 / 0.4328**, which the authors use to support the claim that operation in the RAW domain matters [2607.06217].

## 6. Efficiency, downstream utility, and limitations

Efficiency is a primary part of EeveeDark’s identity. The model is reported at **1.66G FLOPs**, **0.35M parameters**, **~588 ms**, and **0.78 mJ**, compared with **1.49G / 0.30M / ~528 ms / 0.70 mJ** for BRVE, **1.47G / 0.30M / ~521 ms / 0.69 mJ** for BBCU, **24.57G / 10.49M / ~7709 ms / 20.76 mJ** for FloRNN, **32.87G / 13.38M / ~10313 ms / 27.78 mJ** for ShiftNet, and **48.54G / 22.73M / ~15227 ms / 32.57 mJ** for EvLight [2607.06217]. The empirical trade-off claimed by the paper is therefore not maximal quality at any cost, but a favorable quality-efficiency balance relative to both binary baselines and full-precision alternatives.

The paper goes beyond image enhancement metrics and evaluates downstream robotic perception. On **object detection on SDE** using **YOLOv11-s**, mAP@0.5 improves from **0.21** on low-light input to **0.47** with BRVE and **0.73** with EeveeDark [2607.06217]. For **monocular depth estimation**, the paper reports on **SDE**: low light **AbsRel 1.390, RMSE 1.705, \(\delta_1=0.350\), \(\delta_2=0.597\), \(\delta_3=0.768\)**; BRVE **0.921, 1.101, 0.523, 0.785, 0.869**; EeveeDark **0.810, 0.983, 0.568, 0.810, 0.893**. On **CEAR**, low light gives **0.912, 1.860, 0.191, 0.415, 0.585**; BRVE **0.460, 1.230, 0.353, 0.638, 0.812**; and EeveeDark **0.414, 1.060, 0.381, 0.685, 0.863** [2607.06217]. For **visual SLAM on CEAR**, ATE RMSE improves across nearly all sequences; examples given are **around_bldg: 39.93 → 36.36 → 30.56**, **parking_lot1: 40.46 → 11.50 → 6.11**, and **mocap1: 1.77 → 1.50 → 0.27** for low light, BRVE, and EeveeDark respectively [2607.06217].

The paper explicitly presents EeveeDark as **the first end-to-end trainable BNN for low-light video enhancement that jointly processes asynchronous events and sensor-level RAW video** [2607.06217]. In the broader low-light event-vision landscape, this places it in a different niche from Dark-EvGS, which focuses on static-scene radiance field reconstruction and bright novel-view synthesis from dark event and frame data rather than video enhancement [2507.11931]. EeveeDark is therefore a restoration framework rather than a radiance-field method.

The limitations acknowledged in the paper are tied directly to the modality design. When motion is minimal, events become sparse and the event guidance weakens. Under extremely low photon counts, RAW itself is very noisy, which can degrade enhancement quality, produce residual noise, and harm color fidelity [2607.06217]. The system is therefore less effective in static, ultra-dark scenes where both modalities are weak. The most natural deployment settings identified by the paper are edge devices, embedded robotic platforms, mobile or battery-constrained vision systems, and autonomous navigation in low-light environments [2607.06217].

Taken together, EeveeDark is best understood as a multimodal, RAW-domain, efficiency-oriented video restoration framework in which event guidance is used not for explicit event reconstruction, but for temporal modulation and feature refinement inside a predominantly binary architecture. Its contribution lies in the joint integration of sensor-level RAW processing, event-guided low-light video enhancement, and binary neural computation within a single end-to-end trainable system [2607.06217].

Source: https://www.emergentmind.com/topics/eeveedark