Pith. sign in

REVIEW 3 major objections 8 minor 53 references

EeveeDark: A Binary Neural Framework for Low-Light Video Enhancement via Event-Guided Sensor-Level Fusion

T0 review · 3 major / 8 minor · reviewed 2026-07-08 · glm-5.2

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Binary neural net fuses event cameras with RAW sensors for dark video

desk verdict Solid engineering contribution with a real but thin event-guided margin read the letter →

arxiv 2607.06217 v1 pith:A5YAIT6F submitted 2026-07-07 cs.CV

Onur Eker , Erkut Erdem , Aykut Erdem This is my paper
classification cs.CV
keywords eeveedarkbinarylow-lightcomputationaldataenhancementeventevent-guided
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

EeveeDark argues that two sensor modalities — unprocessed Bayer RAW frames and asynchronous event streams — together carry enough complementary information to compensate for the quality loss that normally comes from binarizing a neural network to 1-bit weights and activations. The paper's central claim is that by building modality-specific binary encoders for each input, fusing them in a lightweight block, and using event-driven attention gates to refine temporal features, one can achieve low-light video enhancement quality approaching full-precision models at roughly 5% of their computational cost. The architecture processes RAW frames (which preserve sensor-level color and tonal detail lost in standard RGB) alongside event voxel grids (which capture per-pixel brightness changes at microsecond resolution with high dynamic range). A shift-based encoder propagates temporal context across frames, and an Event-Guided Skip Gate uses distribution statistics from event features to dynamically amplify motion-rich regions while suppressing noise in static areas. The model is trained end-to-end with Charbonnier loss in the RAW domain, outputting enhanced RAW frames that a standard image signal processor then converts to RGB. On the LLRVD dataset, EeveeDark achieves 30.37 PSNR in the RGB domain at 1.66G FLOPs and 0.35M parameters, compared to ShiftNet's 32.01 PSNR at 32.87G FLOPs and 13.38M parameters — roughly a 20× reduction in computation and 38× reduction in parameters for a 1.64 dB quality gap. The paper also evaluates on the real-world HUE dataset (which has no ground truth) using no-reference quality metrics, on SDE and SDSD RGB benchmarks, and on downstream robotic tasks including object detection, monocular depth estimation, and visual SLAM, reporting consistent improvements over prior binary methods.

What carries the argument

Binary Neural Network with distribution-aware binary convolutions; modality-specific binary encoders for RAW frames and event voxel grids; multi-modal feature fusion block combining binary and full-precision convolutions via element-wise addition; cyclic temporal shift mechanism for cross-frame information propagation; Event-Guided Skip Gate (EGSG) using channel-wise distribution statistics (mean, absolute mean, standard deviation) from event features to generate attention maps; spatial shift decoder with predefined shift kernels for motion alignment; selective full-precision layers at input and output to preserve fine spatial detail.

What would settle it

Deploy EeveeDark on a real event-camera-equipped robot in conditions where both photon counts and motion are minimal (the paper's own stated failure regime). If the event-guided gating produces no improvement over the RAW-only baseline in this regime, the central claim that event data compensates for binarization quality loss is conditionally falsified — it would hold only when events are sufficiently dense, not in the general low-light case the paper targets.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the quality gap between binary and full-precision neural networks for low-light video enhancement can be substantially narrowed by introducing event-camera data as a second input modality alongside sensor-level RAW frames, without abandoning the computational efficiency of 1-bit quantization. The Event-Guided Skip Gate mechanism — which extracts channel-wise mean, absolute-mean, and standard-deviation statistics from event features to generate a sigmoid-normalized gating map that modulates temporally shifted features — is the specific architectural component the paper identifies as responsible for translating event-driven motion cues into improved spatiq

Load-bearing premise

The model is trained and primarily evaluated using synthetic event streams generated by the v2e simulator, because no public dataset pairs real event-camera recordings with ground-truth clean RAW video. If the simulated events do not faithfully reproduce the noise, latency, and motion characteristics of physical event sensors, the model's measured performance may not transfer to real-world deployment.

Editorial extensions

If this is right

  • If the performance-efficiency trade-off holds on real hardware, robots and drones operating in low-light conditions could run video enhancement on embedded ARM processors at sub-second latency per frame, enabling real-time perception without GPU acceleration.
  • The downstream task results (object detection mAP of 0.73 vs. 0.21 for low-light inputs, improved SLAM trajectory accuracy) suggest that enhancement quality directly gates robotic autonomy in darkness — the bottleneck is perception, not planning.
  • The RAW-domain processing pipeline demonstrates that operating before the ISP stage preserves dynamic range critical for extreme low-light recovery, which could shift how future edge-vision systems are architected.
  • The event-guided gating mechanism, which uses simple distribution statistics rather than expensive attention or flow estimation, offers a template for incorporating asynchronous sensor data into other efficient architectures.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The model's real-world generalization claims rest on synthetic event simulation (v2e) for training, since no paired RAW-event ground-truth datasets exist. If simulated events systematically differ from real event-camera noise profiles — particularly threshold noise and motion trailing characteristics — the HUE dataset results may not reflect deployment performance. A paired real RAW-event dataset
  • The failure mode the paper identifies (sparse events under minimal motion combined with poor-SNR RAW) suggests that the approach is fundamentally limited by photon scarcity in both modalities simultaneously. This implies a theoretical floor: when neither sensor produces useful signal, no fusion architecture can recover what was never captured.
  • The ~588ms estimated latency on ARM64 is computed from operation-count mappings rather than direct measurement. Actual deployment latency depends on memory bandwidth, cache behavior, and binary-operation kernel optimization, which may not scale linearly with the FLOP reductions reported.
  • The selective full-precision components (first convolution layer, fusion block, output layer, EGSG projections) mean the model is not fully binary — the 0.35M parameter count already accounts for this via the bin/32 accounting, but the ratio of full-precision to binary parameters determines how much of the theoretical efficiency gain is actually realized.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 8 minor

Summary. EeveeDark introduces a Binary Neural Network (BNN) for low-light video enhancement that fuses sensor-level Bayer RAW frames with event streams. The architecture comprises modality-specific binary encoders, a multi-modal fusion block, a shift encoder/decoder with recurrent embeddings, and an Event-Guided Skip Gate (EGSG) that uses event-derived statistics to modulate shifted features. The model is evaluated on LLRVD (RAW domain, synthetic events via v2e), HUE (real events, no ground truth), and SDE/SDSD (RGB benchmarks). The authors also evaluate downstream utility via object detection, depth estimation, and visual SLAM. The paper reports 1.66G FLOPs and 0.35M parameters, positioning EeveeDark between prior BNN baselines (BBCU, BRVE) and full-precision models (ShiftNet, FloRNN, EvLight).

Significance. The paper addresses a practically relevant problem: achieving low-light video enhancement under tight computational budgets suitable for robotic deployment. The integration of event cameras with RAW-level processing in a binarized framework is novel within the BNN literature, and the breadth of evaluation — spanning RAW-domain PSNR, no-reference metrics on real events, RGB benchmarks, and three downstream robotic tasks (detection, depth, SLAM) — is commendable. The inclusion of a project page and the use of established complexity estimation protocols (daBNN, 7nm energy model) add reproducibility value. The favorable performance-efficiency trade-off (Table II: 30.37 PSNR at 1.66G FLOPs vs. ShiftNet's 32.01 at 32.87G) is a meaningful result for the resource-constrained deployment setting the paper targets.

major comments (3)
  1. Table IV (ablation on LLRVD): The entire event-guided contribution amounts to 0.44 PSNR in the RAW domain (0.37 from the Event Encoder, 0.07 from EGSG). These numbers are measured exclusively on v2e-simulated events (Section IV.B). The only real-event evaluation with ground truth is on SDE (Table III), where no ablation is reported and EeveeDark underperforms EvLight by 0.84 dB PSNR. The HUE evaluation (Table II) uses real events but lacks ground truth, and EeveeDark's CLIP-IQA (0.177) and MANIQA (0.158) are below EvLight (0.251, 0.171). The quantitative case that events meaningfully help the BNN thus rests entirely on synthetic data, and the improvement is small enough that it could partly reflect v2e's noise model rather than a genuine event-guided signal. The authors should either (a) run the Table IV ablation on a real-event dataset (even SDE, retrained in the RGB protocol) or (b) re
  2. Section IV.E, HUE results: The text states that EeveeDark 'matches full-precision methods in perceptual quality' on HUE, but Table II shows it trails EvLight on CLIP-IQA and MANIQA while leading only on TOPIQ-NR (0.605 vs. 0.589) and NIQE (5.358 vs. 5.696). The no-reference metrics are mixed, not uniformly favorable. The claim should be softened to reflect that EeveeDark achieves competitive but not superior no-reference quality on HUE, and the discussion should acknowledge that the real-event results do not clearly demonstrate event benefit over the RAW-only BNN baseline.
  3. Table III (SDE): EeveeDark achieves 21.99 PSNR vs. EvLight's 22.83, a gap of 0.84 dB, while using ~29x fewer FLOPs (1.66G vs. 48.54G). This is a legitimate efficiency-quality trade-off, but the paper's framing in the introduction and conclusion ('favorable performance-efficiency trade-off') does not adequately discuss that on the only real-event benchmark with ground truth, the model underperforms the primary full-precision event-guided baseline. A more balanced discussion of this trade-off, including whether the gap is acceptable for the targeted robotic deployment scenarios, would strengthen the contribution.
minor comments (8)
  1. Section III.A: The RAW frame preprocessing description could be clearer. The scaling factor r is mentioned but its value or selection criterion is not specified. Stating the value used in experiments would aid reproducibility.
  2. Section III.B, Eq. (3): The fusion block combines binary and full-precision convolutions via element-wise addition. The motivation for this specific combination (rather than, e.g., concatenation followed by a single convolution) is not discussed. A brief justification would help.
  3. Section III.B, Eq. (7): The channel-wise statistics (mean, absolute mean, std) are concatenated to form Z_t. The dimensionality of Z_t and the number of channels in the attention convolution should be specified.
  4. Section II: REN [21] and EvLight++ [22] are mentioned as follow-up works improving temporal consistency, but no quantitative comparison is provided. Given that EvLight++ is published in TPAMI 2025, at least a brief discussion of how EeveeDark's approach differs or a note that direct comparison was not feasible would be appropriate.
  5. Section IV.C: The FLOPs estimation formula (FLOPs = OPs_bin/64 + OPs_fp) is standard but the resolution at which OPs are counted (256x256) should be cross-referenced more explicitly with the per-frame numbers in Table II to avoid confusion.
  6. Table V: The SSIM value for RAW2RAW+ISP is reported as 0.8501 with four decimal places, while other entries in Tables I and II use three. Consistency in decimal places would be cleaner.
  7. Fig. 2: The data flow from the fusion block to the shift encoder is somewhat difficult to follow. The EGSG blocks are shown but their connection to the event encoder features is not clearly drawn. A revised figure with clearer arrows or a textual walkthrough would improve understanding.
  8. Section IV.H (Limitations): The acknowledgment that performance degrades 'under extremely low photon counts and minimal motion' is honest and appreciated. This could be strengthened by quantifying the degradation (e.g., at what lux level or event density does performance drop sharply).

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for a careful and constructive review. The referee raises three interconnected concerns: (1) the event-guided contribution is validated primarily on synthetic (v2e) events with a modest 0.44 PSNR gain, (2) the HUE no-reference metrics are mixed rather than uniformly favorable, and (3) the SDE gap to EvLight (0.84 dB) is under-discussed relative to the 'favorable trade-off' framing. We agree with all three points and will revise the manuscript accordingly. Specifically, we will (a) add an ablation on SDE retrained in the RGB protocol to assess event contribution on real events, (b) soften the HUE claims to 'competitive' rather than 'matching,' and (c) add a balanced discussion of the SDE–EvLight gap in the introduction and conclusion. We also note that the 0.44 PSNR ablation gain, while modest in absolute terms, is consistent with the magnitude of improvements reported by prior event-guided methods and is accompanied by larger gains in temporal consistency (ST-RRED: 0.046 → 0.039) and downstream task performance (e.g., depth estimation AbsRel: 0.921 → 0.810 on SDE), which we will make more prominent in the revision.

read point-by-point responses

  1. Referee: Table IV ablation rests entirely on v2e-simulated events; the 0.44 PSNR gain could reflect v2e's noise model rather than genuine event-guided signal. The referee requests either (a) running the ablation on a real-event dataset or (b) re-discussing the limitation.

    Authors: The referee is correct that the Table IV ablation is conducted exclusively on v2e-simulated events and that the 0.44 PSNR gain is modest. We will address this by running the ablation (w/o Event Encoder, w/o EGSG, full model) on SDE retrained under the RGB protocol described in Section IV.B. SDE uses real events from a DAVIS346 sensor, so this will directly test whether the event-guided contribution holds on real event data. We will report these results in a revised Table IV (or an additional table). We also agree that the current manuscript does not adequately acknowledge the synthetic-event limitation of the ablation, and we will add an explicit discussion of this point. We would note, however, that the event contribution is not captured by PSNR alone: the ST-RRED improvement (0.046 → 0.039, a 15% relative reduction) indicates a more substantial gain in temporal stability, and the downstream task results on SDE (which uses real events) show larger practical gains — e.g., depth estimation AbsRel improves from 0.921 (BRVE, no events) to 0.810 (EeveeDark, with events), and object detection mAP improves from 0.47 to 0.73. These downstream results on real-event data provide corroborating evidence that the event-guided signal is genuine and not merely an artifact of v2e's noise model. We will make this argument explicit in the revision. revision: yes

  2. Referee: Section IV.E claims EeveeDark 'matches full-precision methods in perceptual quality' on HUE, but Table II shows mixed no-reference metrics: EvLight leads on CLIP-IQA (0.251 vs. 0.177) and MANIQA (0.171 vs. 0.158), while EeveeDark leads on TOPIQ-NR (0.605 vs. 0.589) and NIQE (5.358 vs. 5.696). The claim should be softened.

    Authors: We agree. The claim that EeveeDark 'matches full-precision methods in perceptual quality' on HUE is not supported by the full set of no-reference metrics. As the referee notes, the results are mixed: EeveeDark leads on TOPIQ-NR and NIQE but trails EvLight on CLIP-IQA and MANIQA. We will revise the text in Section IV.E to state that EeveeDark achieves 'competitive but not uniformly superior' no-reference quality on HUE, and we will add a sentence acknowledging that the real-event results do not clearly demonstrate event benefit over the RAW-only BNN baseline on these particular metrics. We believe the mixed results are themselves informative — they suggest that the event-guided contribution is most visible in temporal stability and downstream task performance rather than in single-frame perceptual metrics — and we will add this interpretation to the discussion. revision: yes

  3. Referee: On SDE (the only real-event benchmark with ground truth), EeveeDark underperforms EvLight by 0.84 dB PSNR. The 'favorable performance-efficiency trade-off' framing in the introduction and conclusion does not adequately discuss this gap.

    Authors: We agree that the manuscript's framing should more directly acknowledge the SDE gap. We will revise the introduction and conclusion to state explicitly that on the only real-event benchmark with ground truth (SDE), EeveeDark underperforms the primary full-precision event-guided baseline (EvLight) by 0.84 dB PSNR, while using approximately 29× fewer FLOPs (1.66G vs. 48.54G). We will add a discussion of whether this gap is acceptable for the targeted robotic deployment scenarios, noting that: (i) EeveeDark outperforms all image-only methods on SDE (e.g., Retinexformer at 22.11 PSNR vs. our 21.99 is within noise, but Uformer at 21.71 and SNR-Net at 21.12 are clearly below), (ii) the downstream task results on SDE show that the PSNR gap does not translate to a proportional degradation in task-level utility — EeveeDark's depth estimation (AbsRel 0.810) and object detection (mAP 0.73) substantially outperform BRVE, suggesting the enhanced frames remain practically useful, and (iii) the 29× computational reduction is the central design constraint, and the gap represents the quality cost of binarization. We will present this as an honest trade-off discussion rather than a purely favorable characterization. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: empirical systems paper with self-contained evaluation against external benchmarks

full rationale

This is an empirical systems paper, not a derivation chain. The central claim — that EeveeDark achieves a favorable performance-efficiency trade-off as a BNN for low-light video enhancement — is evaluated against external benchmarks (LLRVD, SDE, SDSD, HUE) and compared against external baselines (BBCU, BRVE, ShiftNet, FloRNN, EvLight). Architectural components (distribution-aware binary convolutions [9], cyclic temporal shift [10], RPReLU [27], v2e event simulation [30]) are cited from prior work without claiming them as novel derivations or predictions. The ablation study (Table IV) measures module contributions empirically, not by definition. The loss function (Charbonnier) is standard. No step in the paper reduces to its inputs by construction, no prediction is a renamed fit, and no self-citation is load-bearing for a mathematical claim. The concerns raised by the skeptic (synthetic events, marginal event contribution) are correctness and generalization risks, not circularity. The paper is self-contained against external benchmarks.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The paper does not invent new mathematical entities, particles, or physical forces. It introduces an engineering architecture (EeveeDark) composed of known components (binary convolutions, temporal shift, event gating). The free parameters are standard deep learning hyperparameters. The axioms are domain assumptions about the effectiveness of prior work components and simulation tools.

free parameters (5)
  • B (temporal bins for event voxel grids) = 5
    Stated in Section III.A: 'We use B=5 temporal bins in all experiments.' This is a hyperparameter chosen empirically.
  • r (RAW amplification scaling factor)
    Mentioned in Section III.A as part of preprocessing; specific value not stated but is a data-dependent parameter.
  • Learning rate schedule = 2e-4 to 1e-7
    Section IV.A specifies cosine annealing restart from 2×10^-4 to 1×10^-7 for 100k iterations.
  • Spatial shift kernel K = {(x,y)|x,y in {-8,-4,0,4,8}, (x,y)!=(0,0)}
    Section III.B, Shift Decoder: predefined shift kernel expanding receptive field.
  • Charbonnier loss epsilon = 1e-3
    Section III.B, Loss Function: epsilon=1×10^-3 for numerical stability.
assumptions (5)
  • domain assumption Distribution-aware binary convolutions [9] effectively reduce computational complexity while maintaining feature quality in low-light video enhancement.
    Section III.B: The encoders employ distribution-aware binary convolutions from [9] (BRVE). This is a foundational building block assumed to work as claimed.
  • domain assumption Cyclic temporal shift mechanism [10] effectively propagates long-range temporal information across frames.
    Section III.B, Shift Encoder: The cyclic temporal shift mechanism from [10] (ShiftNet) is used for temporal information exchange.
  • domain assumption Synthetic events generated by v2e [30] (improved version from [31]) are sufficiently realistic to train models that generalize to real event cameras.
    Section IV.B: LLRVD training uses synthetic events from v2e. The model's performance claims on the primary benchmark depend on this assumption.
  • domain assumption daBNN framework [43] provides an accurate empirical mapping between operation counts and execution time on ARM64 CPUs.
    Section IV.C: Latency estimation relies on daBNN. The efficiency claims are estimated, not directly measured.
  • domain assumption The 7nm device model from [44] accurately estimates energy consumption for BNN operations.
    Section IV.C: Energy consumption is estimated according to this model. Used to support efficiency claims.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EeveeDark: A Binary Neural Framework for Low-Light Video Enhancement via Event-Guided Sensor-Level Fusion." pith.science (2026). https://pith.science/paper/A5YAIT6F

@misc{pith2026260706217,
  author       = {Pith},
  title        = {Pith review of: EeveeDark: A Binary Neural Framework for Low-Light Video Enhancement via Event-Guided Sensor-Level Fusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A5YAIT6F}},
  note         = {Machine review of arXiv:2607.06217}
}
read the original abstract

Enhancing videos under extreme low-light conditions remains challenging due to the difficulty of balancing restoration quality and computational efficiency in resource-constrained settings. This paper introduces EeveeDark, a low-light video enhancement framework that combines the spatial richness of sensor-level RAW data with the temporal precision of event streams. Central to our model is a Binary Neural Network (BNN) architecture that reduces computational overhead by quantizing weights and activations while preserving detail. EeveeDark incorporates (i) modality-specific binary encoders for processing RAW frames and event data, (ii) a lightweight fusion block for integrating spatial and temporal cues, and (iii) an event-guided skip gating mechanism for dynamic spatiotemporal refinement. Experiments on synthetic and real-world datasets show that EeveeDark outperforms prior BNN-based methods and offers a favorable performance-efficiency trade-off compared to full-precision models. The project page is available at https://cyberiada.github.io/EeveeDark.

Figures

Figures reproduced from arXiv: 2607.06217 by the authors.

Figure 1
Figure 1. Comparison with state-of-the-art models. PSNR versus compu￾tational complexity is shown, with circle sizes indicating parameter count (millions). EeveeDark outperforms prior BNN methods BBCU and BRVE by a significant margin, and offers a favorable performance-efficiency trade-off against full-precision models, ShiftNet, FloRNN, and EvLight. discards sensor-level information critical for faithful color and tonal rest… view at source ↗
Figure 2
Figure 2. Overview of EeveeDark. It comprises five key components: (1) Preprocessing and modality-specific encoding, where Bayer RAW frames and event data are separately encoded using binary encoders; (2) an Efficient Multi-Modal Fusion Block that integrates the spatial detail of RAW inputs with the temporal precision of event streams; (3) a Shift Encoder that propagates temporal information across frames and modulates featur… view at source ↗
Figure 3
Figure 3. Event-Guided Skip Gate (EGSG). It combines the temporal precision of event data with the spatial context of shifted features. Event voxel features are interpolated and projected to the encoder’s resolution to generate an attention map that modulates shifted features, emphasizing dynamic regions and suppressing redundant information. B. Datasets LLRVD (RAW Training with Synthetic Events). We eval￾uate RAW-domain perf… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison on the LLRVD dataset. EeveeDark gen￾erates clean, detailed outputs with strong temporal consistency, outperforming binary baselines (BBCU, BRVE) and approaching the visual quality of full￾precision models like ShiftNet and FloRNN. Unlike EvLight,…
Figure 5
Figure 5. Figure 5: Qualitative comparison on the real-world HUE dataset. EeveeDark generalizes well, avoiding the oversmoothing of BNN-based methods (BBCU, BRVE) and the static artifacts of EvLight. Unlike full-precision models (FloRNN, ShiftNet), which show color distortions and grain, …
Figure 6
Figure 6. Figure 6: Qualitative comparisons on SDE and SDSD datasets. BRVE often blurs fine structures and amplifies noise in dark regions, whereas EeveeDark produces sharper edges and improved local contrast by effectively leveraging event information. Best viewed with zoom. BRVE EeveeDa…
Figure 7
Figure 7. Figure 7: Temporal profile comparisons. The temporal profile is generated by stacking a vertical line (shown in yellow) across consecutive frames. BRVE shows flickering and instability over time, whereas EeveeDark yields a smoother, temporally consistent output. Best viewed when…
Figure 8
Figure 8. Figure 8: Comparison on downstream tasks. EeveeDark improves object detection and depth estimation under dark conditions compared to BRVE. TABLE VII VISUAL SLAM EVALUATION ON CEAR DATASET. WE REPORT THE ABSOLUTE TRAJECTORY ERROR (ATE) RMSE [M]. Sequence L.Light BRVE Ours Sequenc…
Figure 9
Figure 9. Figure 9: Visual analysis of limitations and failure cases. The examples illustrate two primary challenges: Performance degradation in static scenes due to sparse event guidance, and artifacts in extreme low-light regions. shown in [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 53 canonical work pages

  1. [1]

    Low-light image and video enhancement using deep learning: A survey,

    C. Li, C. Guo, L. Han, J. Jiang, M.-M. Cheng, J. Gu, and C. C. Loy, “Low-light image and video enhancement using deep learning: A survey,”IEEE Trans. Pattern Anal. Mach., vol. 44, no. 12, pp. 9396– 9416, 2022. 8 IEEE ROBOTICS AND AUTOMATION LETTERS. PREPRINT VERSION. ACCEPTED FEBRUARY , 2026

    work page 2022

  2. [2]

    Towards robust event- guided low-light image enhancement: a large-scale real-world event- image dataset and novel approach,

    G. Liang, K. Chen, H. Li, Y . Lu, and L. Wang, “Towards robust event- guided low-light image enhancement: a large-scale real-world event- image dataset and novel approach,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 23– 33

    work page 2024

  3. [3]

    Coherent event guided low-light video enhancement,

    J. Liang, Y . Yang, B. Li, P. Duan, Y . Xu, and B. Shi, “Coherent event guided low-light video enhancement,” inProc. ICCV, 2023, pp. 10 615– 10 625

    work page 2023

  4. [4]

    Low-light video enhancement with synthetic event guidance,

    L. Liu, J. An, J. Liu, S. Yuan, X. Chen, W. Zhou, H. Li, Y . F. Wang, and Q. Tian, “Low-light video enhancement with synthetic event guidance,” inProc. AAAI, vol. 37, no. 2, 2023, pp. 1692–1700

    work page 2023

  5. [5]

    Event-based low-illumination image enhancement,

    Y . Jiang, Y . Wang, S. Li, Y . Zhang, M. Zhao, and Y . Gao, “Event-based low-illumination image enhancement,”IEEE Trans. Multimedia, 2023

    work page 2023

  6. [6]

    Binarized neural networks,

    I. Hubara, M. Courbariaux, D. Soudry, R. El-Yaniv, and Y . Bengio, “Binarized neural networks,”Advances in neural information processing systems, vol. 29, 2016

    work page 2016

  7. [7]

    Xnor-net: Imagenet classification using binary convolutional neural networks,

    M. Rastegari, V . Ordonez, J. Redmon, and A. Farhadi, “Xnor-net: Imagenet classification using binary convolutional neural networks,” in Proc. ECCV, 2016, pp. 525–542

    work page 2016

  8. [8]

    Basic binary convolution unit for binarized image restoration network,

    B. Xia, Y . Zhang, Y . Wang, Y . Tian, W. Yang, R. Timofte, and L. Van Gool, “Basic binary convolution unit for binarized image restoration network,” inProc. ICLR, 2023

    work page 2023

Show all 53 references
  1. [9]

    Binarized low-light RAW video enhancement,

    G. Zhang, Y . Zhang, X. Yuan, and Y . Fu, “Binarized low-light RAW video enhancement,” inProc. CVPR, 2024, pp. 25 753–25 762

    2024

  2. [10]

    A simple baseline for video restoration with grouped spatial- temporal shift,

    D. Li, X. Shi, Y . Zhang, K. C. Cheung, S. See, X. Wang, H. Qin, and H. Li, “A simple baseline for video restoration with grouped spatial- temporal shift,” inProc. CVPR, 2023, pp. 9822–9832

    2023

  3. [11]

    MBLLEN: Low-light image/video enhancement using cnns

    F. Lv, F. Lu, J. Wu, and C. Lim, “MBLLEN: Low-light image/video enhancement using cnns.” inProc. BMVC, 2018

    2018

  4. [12]

    Zero- reference deep curve estimation for low-light image enhancement,

    C. Guo, C. Li, J. Guo, C. C. Loy, J. Hou, S. Kwong, and R. Cong, “Zero- reference deep curve estimation for low-light image enhancement,” in Proc. CVPR, 2020, pp. 1780–1789

    2020

  5. [13]

    Learning to see in the dark,

    C. Chen, Q. Chen, J. Xu, and V . Koltun, “Learning to see in the dark,” inProc. CVPR, 2018, pp. 3291–3300

    2018

  6. [14]

    Uformer: A general u-shaped transformer for image restoration,

    Z. Wang, X. Cun, J. Bao, W. Zhou, J. Liu, and H. Li, “Uformer: A general u-shaped transformer for image restoration,” inProc. CVPR, 2022, pp. 17 683–17 693

    2022

  7. [15]

    Retinex- former: One-stage Retinex-based transformer for low-light image en- hancement,

    Y . Cai, H. Bian, J. Lin, H. Wang, R. Timofte, and Y . Zhang, “Retinex- former: One-stage Retinex-based transformer for low-light image en- hancement,” inProc. ICCV, 2023, pp. 12 504–12 513

    2023

  8. [16]

    Diffuseraw: End-to-end generative RAW image processing for low-light images,

    R. Dagli, “Diffuseraw: End-to-end generative RAW image processing for low-light images,”arXiv preprint arXiv:2402.18575, 2023

    2023 arXiv

  9. [17]

    Dvdnet: A fast network for deep video denoising,

    M. Tassano, J. Delon, and T. Veit, “Dvdnet: A fast network for deep video denoising,” inProc. ICIP, 2019, pp. 1805–1809

    2019

  10. [18]

    Unidirectional video denoising by mimicking backward recurrent modules with look-ahead forward ones,

    J. Li, X. Wu, Z. Niu, and W. Zuo, “Unidirectional video denoising by mimicking backward recurrent modules with look-ahead forward ones,” inProc. ECCV, 2022, pp. 592–609

    2022

  11. [19]

    Learning to see in the dark with events,

    S. Zhang, Y . Zhang, Z. Jiang, D. Zou, J. Ren, and B. Zhou, “Learning to see in the dark with events,” inProc. ECCV, 2020, pp. 666–682

    2020

  12. [20]

    Event enhanced high- quality image recovery,

    B. Wang, J. He, L. Yu, G.-S. Xia, and W. Yang, “Event enhanced high- quality image recovery,” inProc. ECCV, 2020, pp. 155–171

    2020

  13. [21]

    Exploring in extremely dark: Low-light video enhancement with real events,

    X. Wang, H. Fu, J. Wang, X. Wang, H. Zhang, and H. Ma, “Exploring in extremely dark: Low-light video enhancement with real events,” in Proc. ICME, 2024, pp. 4805–4813

    2024

  14. [22]

    Evlight++: Low- light video enhancement with an event camera: A large-scale real- world dataset, novel method, and more,

    K. Chen, G. Liang, Y . Lu, H. Li, and L. Wang, “Evlight++: Low- light video enhancement with an event camera: A large-scale real- world dataset, novel method, and more,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025

  15. [23]

    Rgb-event isp: The dataset and benchmark,

    L. Yunfan, Y . Qian, Z. Rao, J. Xiao, L. Chen, and H. Xiong, “Rgb-event isp: The dataset and benchmark,” inProc. ICLR, 2024

    2024

  16. [24]

    Binary neural networks: A survey,

    H. Qin, R. Gong, X. Liu, X. Bai, J. Song, and N. Sebe, “Binary neural networks: A survey,”Pattern Recognition, vol. 105, p. 107281, 2020

    2020

  17. [25]

    Binarized Mamba-transformer for lightweight quad Bayer HybridEVS demosaicing,

    S. Zhou, H. Zeng, Y . Lu, T. Shao, K. Tang, Y . Chen, J. Liu, and J. Su, “Binarized Mamba-transformer for lightweight quad Bayer HybridEVS demosaicing,” inProc. CVPR, 2025

    2025

  18. [26]

    Events-to-video: Bringing modern computer vision to event cameras,

    H. Rebecq, R. Ranftl, V . Koltun, and D. Scaramuzza, “Events-to-video: Bringing modern computer vision to event cameras,” inProc. CVPR, 2019, pp. 3857–3866

    2019

  19. [27]

    Reactnet: Towards precise binary neural network with generalized activation functions,

    Z. Liu, Z. Shen, M. Savvides, and K.-T. Cheng, “Reactnet: Towards precise binary neural network with generalized activation functions,” in Proc. ECCV, 2020, pp. 143–159

    2020

  20. [28]

    Deep laplacian pyramid networks for fast and accurate super-resolution,

    W.-S. Lai, J.-B. Huang, N. Ahuja, and M.-H. Yang, “Deep laplacian pyramid networks for fast and accurate super-resolution,” inProc. CVPR, 2017

    2017

  21. [29]

    Low-light RAW video denoising with a high-quality realistic motion dataset,

    Y . Fu, Z. Wang, T. Zhang, and J. Zhang, “Low-light RAW video denoising with a high-quality realistic motion dataset,”IEEE Trans. Multimedia, vol. 25, pp. 8119–8131, 2022

    2022

  22. [30]

    v2e: From video frames to realistic dvs events,

    Y . Hu, S.-C. Liu, and T. Delbruck, “v2e: From video frames to realistic dvs events,” inProc., 2021, pp. 1312–1321

    2021

  23. [31]

    From sim-to-real: Toward general event-based low-light frame interpolation with per-scene optimization,

    Z. Zhang, Y . Ma, Y . Chen, F. Zhang, J. Gu, T. Xue, and S. Guo, “From sim-to-real: Toward general event-based low-light frame interpolation with per-scene optimization,” inProc. SIGGRAPH Asia, 2024

    2024

  24. [32]

    Real-time intermediate flow estimation for video frame interpolation,

    Z. Huang, T. Zhang, W. Heng, B. Shi, and S. Zhou, “Real-time intermediate flow estimation for video frame interpolation,” inProc. ECCV, 2022, pp. 624–642

    2022

  25. [33]

    Hue dataset: High- resolution event and frame sequences for low-light vision,

    B. Ercan, O. Eker, A. Erdem, and E. Erdem, “Hue dataset: High- resolution event and frame sequences for low-light vision,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 174–191

    2024

  26. [34]

    Seeing dynamic scene in the dark: A high-quality video dataset with mechatronic alignment,

    R. Wang, X. Xu, C.-W. Fu, J. Lu, B. Yu, and J. Jia, “Seeing dynamic scene in the dark: A high-quality video dataset with mechatronic alignment,” inProc. ICCV, 2021, pp. 9700–9709

    2021

  27. [35]

    Image quality assessment: from error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,”IEEE Trans. Image Process., vol. 13, no. 4, pp. 600–612, 2004

    2004

  28. [36]

    Video quality assessment by reduced reference spatio-temporal entropic differencing,

    R. Soundararajan and A. C. Bovik, “Video quality assessment by reduced reference spatio-temporal entropic differencing,”IEEE Trans. Circuits Syst. Video Technol., vol. 23, no. 4, pp. 684–694, 2012

    2012

  29. [37]

    Exploring clip for assessing the look and feel of images,

    J. Wang, K. C. Chan, and C. C. Loy, “Exploring clip for assessing the look and feel of images,” inProc. AAAI, 2023

    2023

  30. [38]

    Maniqa: Multi-dimension attention network for no-reference image quality assessment,

    S. Yang, T. Wu, S. Shi, S. Lao, Y . Gong, M. Cao, J. Wang, and Y . Yang, “Maniqa: Multi-dimension attention network for no-reference image quality assessment,” inProc. CVPR, 2022, pp. 1191–1200

    2022

  31. [39]

    Topiq: A top-down approach from semantics to distortions for image quality assessment,

    C. Chen, J. Mo, J. Hou, H. Wu, L. Liao, W. Sun, Q. Yan, and W. Lin, “Topiq: A top-down approach from semantics to distortions for image quality assessment,”IEEE Transactions on Image Processing, vol. 33, pp. 2404–2418, 2024

    2024

  32. [40]

    Making a “completely blind

    A. Mittal, R. Soundararajan, and A. C. Bovik, “Making a “completely blind” image quality analyzer,”IEEE Signal Process. Lett., vol. 20, no. 3, pp. 209–212, 2013

    2013

  33. [41]

    Binarized spectral compressive imaging,

    Y . Cai, Y . Zheng, J. Lin, X. Yuan, Y . Zhang, and H. Wang, “Binarized spectral compressive imaging,” inProc. NeurIPS, 2024

    2024

  34. [42]

    Bi- narized diffusion model for image super-resolution,

    Z. Chen, H. Qin, Y . Guo, X. Su, X. Yuan, L. Kong, and Y . Zhang, “Bi- narized diffusion model for image super-resolution,” inProc. NeurIPS, 2024, pp. 30 651–30 669

    2024

  35. [43]

    DABNN: A super fast inference framework for binary neural networks on arm devices,

    J. Zhang, Y . Pan, T. Yao, H. Zhao, and T. Mei, “DABNN: A super fast inference framework for binary neural networks on arm devices,” inProc. ICME, 2019, pp. 2272–2275

    2019

  36. [44]

    PokeBNN: A binary pursuit of lightweight accuracy,

    Y . Zhang, Z. Zhang, and L. Lew, “PokeBNN: A binary pursuit of lightweight accuracy,” inProc. CVPR, 2022, pp. 12 475–12 485

    2022

  37. [45]

    Bitnet: Scaling 1-bit transformers for large language models,

    H. Wang, S. Ma, L. Dong, S. Huang, H. Wang, L. Ma, F. Yang, R. Wang, Y . Wu, and F. Wei, “Bitnet: Scaling 1-bit transformers for large language models,”arXiv preprint arXiv:2310.11453, 2023

    2023 arXiv

  38. [46]

    SNR-aware low-light image enhancement,

    X. Xu, R. Wang, C.-W. Fu, and J. Jia, “SNR-aware low-light image enhancement,” inProc. CVPR, 2022, pp. 17 714–17 724

    2022

  39. [47]

    Reducing the sim-to-real gap for event cameras,

    T. Stoffregen, C. Scheerlinck, D. Scaramuzza, T. Drummond, N. Barnes, L. Kleeman, and R. Mahony, “Reducing the sim-to-real gap for event cameras,” inProc. ECCV, 2020, pp. 534–549

    2020

  40. [48]

    Cear: Comprehensive event camera dataset for rapid perception of agile quadruped robots,

    S. Zhu, Z. Xiong, and D. Kim, “Cear: Comprehensive event camera dataset for rapid perception of agile quadruped robots,”IEEE Robotics and Automation Letters, 2024

    2024

  41. [49]

    Yolov11: An overview of the key architectural enhancements,

    R. Khanam and M. Hussain, “Yolov11: An overview of the key architectural enhancements,”arXiv preprint arXiv:2410.17725, 2024

    2024 arXiv

  42. [50]

    The pascal visual object classes (voc) challenge,

    M. Everingham, L. Van Gool, C. K. Williams, J. Winn, and A. Zisser- man, “The pascal visual object classes (voc) challenge,”International journal of computer vision, vol. 88, no. 2, pp. 303–338, 2010

    2010

  43. [51]

    Depth anything v2,

    L. Yang, B. Kang, Z. Huang, Z. Zhao, X. Xu, J. Feng, and H. Zhao, “Depth anything v2,” inProc. NeurIPS, 2024, pp. 21 875–21 911

    2024

  44. [52]

    Orb-slam3: An accurate open-source library for visual, visual– inertial, and multimap slam,

    C. Campos, R. Elvira, J. J. G. Rodr ´ıguez, J. M. Montiel, and J. D. Tard´os, “Orb-slam3: An accurate open-source library for visual, visual– inertial, and multimap slam,”IEEE transactions on robotics, vol. 37, no. 6, pp. 1874–1890, 2021

    2021

  45. [53]

    evo: Python package for the evaluation of odometry and slam

    M. Grupp, “evo: Python package for the evaluation of odometry and slam.” https://github.com/MichaelGrupp/evo, 2017

    2017

Pith tools

Reviewed July 8, 2026 · model on record in the stance chip above.