Endo-4DGX: Robust Endoscopic Scene Reconstruction and Illumination Correction with Gaussian Splatting

Beilei Cui; Hongbin Liu; Hongliang Ren; Jie Wang; Jinlin Wu; Long Bai; Tong Chen; Yanheng Li; Yiming Huang; Zhen Lei

REVIEW 3 major objections 6 minor 32 references

Endo-4DGX: Robust Endoscopic Scene Reconstruction and Illumination Correction with Gaussian Splatting

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Endo-4DGX adds per-view illumination embeddings to 4D Gaussian Splatting, achieving exposure-corrected endoscopic reconstruction that beats reconstruction-plus-restoration baselines by over 6 dB PSNR.

desk verdict Useful engineering integration of known splatting/illumination tricks, but the headline correction numbers rest on a test protocol that feeds in a training-frame embedding, so the 6 dB gains are likely over-optimistic. read the letter →

arxiv 2506.23308 v1 pith:WXDQ5XW6 submitted 2025-06-29 cs.CV

Yiming Huang , Long Bai , Beilei Cui , Yanheng Li , Tong Chen , Jie Wang , Jinlin Wu , Zhen Lei

show 2 more authors

Hongbin Liu Hongliang Ren

This is my paper · ORCID

classification cs.CV

keywords 3DGaussianSplatting4Dendoscopicscenereconstructionilluminationcorrectionexposurecontrolroboticsurgerydynamicrendering

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

Endo-4DGX extends 4D Gaussian Splatting so a single dynamic scene model can both reconstruct deformable endoscopic tissue and correct uneven lighting in the rendered frames. The paper argues that adding a per-view illumination embedding, two lightweight networks that adjust Gaussian color first locally and then globally, and a global exposure-control loss lets the optimization survive low-light and over-exposed training frames instead of collapsing. On the EndoNeRF exposure-correction dataset the method reports roughly 34–35 dB PSNR with corrected renderings, more than 6 dB above the strongest reconstruction-plus-restoration baseline, while keeping real-time inference at 61 FPS. The motivation is surgical: if the method works in practice, a single model could feed a robot-assisted surgery system with illuminated, geometry-consistent views of the operating field.

What carries the argument

The load-bearing mechanism is the illumination-adaptive Gaussian rendering pipeline: a trainable illumination embedding $\mathbf{e}$ per input view (dimension 32), a region-aware enhancement network $f_{\text{region}}$ that predicts per-Gaussian affine brightness parameters $(\beta, \gamma)$ separately for bright and dark frames, a spatial-aware adjustment network $f_{\text{spatial}}$ producing a view-consistent quadratic curve $\delta$ so that $\tilde{C} = C_{\text{tone}} + \delta C_{\text{tone}}(1 - C_{\text{tone}})$, and a global exposure control loss that pulls the mean rendered color toward a fixed exposure level (0.6). The illumination embedding carries view-dependent brightness information through training; the two adjustment modules refine illumination at Gaussian level and image level respectively; and the exposure loss stabilizes optimization against dark or bright inputs and enables correction at test time using a normal-exposure frame embedding.

What would settle it

Take a real endoscopic video with paired ground-truth normal-exposure frames, apply Endo-4DGX exactly as described, and compare corrected PSNR against a 2D restoration network plus 4DGS baseline; if the gain drops below statistical significance, the synthetic-exposure protocol is the reason. A cheaper check: ablate the test-time embedding by estimating a per-frame embedding from the test input itself; if quality collapses, the correction depends on access to a clean training frame's embedding.

Watch

Extended reading notes

Core claim

The paper's central claim is that illumination-adaptive Gaussian Splatting, implemented as trainable per-frame illumination embeddings $\mathbf{e} \in \mathbb{R}^{N \times 32}$ jointly optimized with 4D Gaussians, plus a region-aware enhancement network that applies $c_{\text{tone}} = \beta c + \gamma$ separately for bright and dark frames, plus a spatial-aware adjustment $\tilde{C}_{\text{tone}} = C_{\text{tone}} + \delta C_{\text{tone}}(1 - C_{\text{tone}})$, plus a global exposure control loss, yields state-of-the-art rendering quality under both low-light and over-exposure while preserving geometric accuracy. The claim is tested against combinations of 2D restoration methods (EndoUIC, CSEC) with deformable Gaussian baselines (Deform3DGS, Endo-4DGS, EndoGaussian) and against general-scene illumination Gaussian methods (WildGaussians, Gaussian-DK, DarkGS), with the proposed method reporting the highest PSNR/SSIM and lowest LPIPS on all datasets. The authors further claim the correction works without a normal-exposure embedding in the scene by relying on the exposure control loss, and that the method generalizes to uneven-illumination clips from StereoMIS and C3VD.

Load-bearing premise

The reported gains assume that the EndoNeRF-EC test protocol—synthetically varying exposure with Adobe Camera Raw and correcting test frames using the illumination embedding of the first normal-exposure training frame—faithfully represents how a surgeon's lighting changes in real endoscopic video, so that the 6 dB improvement would transfer to actual surgical conditions.

Editorial extensions

If this is right

A single Endo-4DGX model can both reconstruct a deformable surgical scene and output exposure-corrected renderings in real time, removing the need to cascade a separate 2D restoration network before reconstruction.
GPU memory consumption is lower than the 2D-restoration-plus-4DGS pipeline (about 46% reduction reported), making the approach more feasible on clinical hardware.
The method handles scenes without any normal-exposure training frames via the global exposure control loss, broadening applicability to endoscope videos that are consistently under- or over-exposed.
The reported generalization to StereoMIS and C3VD clips suggests the illumination-adaptive components transfer across surgical domains (prostatectomy, porcine surgery, cardiovascular colonoscopy) without retraining.
The classification rule $IC(I) = \text{mean}(I) > \text{mean}(p)$ provides a simple, training-free routing of frames to dark versus bright enhancement networks.

Reading between the lines

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

The paper's test-time correction uses the embedding of the first normal-exposure training frame; a natural extension the authors do not develop is estimating the embedding directly from the test frame itself, which would turn the method into a true burst-exposure correction system.
Because the region-aware and spatial-aware modules are lightweight MLPs, the same illumination-adaptive machinery could be ported to other deformable radiance field backbones, not only 4DGS.
The synthetic exposure protocol may underestimate real-world failure modes: real endoscope light sources change color temperature and create specular reflections, so the 6 dB claim should be re-tested on paired real footage before relying on it for surgical navigation.
The global exposure control loss's target level $E = 0.6$ is a fixed scalar; adapting $E$ per scene or per tissue type could improve results on datasets with different brightness distributions.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The paper does something genuinely useful: it takes per-frame illumination embeddings (WildGaussians), affine and quadratic color adjustments (ALeth-NeRF, Zero-DCE), and bolts them onto a 4DGS backbone for endoscopic scenes. The ablations show each piece matters, the reported FPS and memory numbers are sensible, and the idea of routing dark vs bright frames through separate small MLPs is a reasonable practical choice. If the evaluation were trustworthy, this would be a solid contribution to surgical scene reconstruction.

But the evaluation has a load-bearing flaw that the stress-test note catches correctly. For EndoNeRF-EC correction, Section 3.2 says they use the first training frame with normal illumination as the input embedding. That embedding is a trainable per-frame parameter, optimized jointly with the Gaussians and supervised by the color loss against that very frame's normal appearance. Feeding that embedding at test time means the correction networks are not seeing a representation of the test image's lighting; they see an index for a training frame whose normal appearance they were explicitly trained to reproduce. Since the test frames come from the same video clips and are temporally close to that training frame, the 6.15 and 10.76 dB gains over EndoUIC+Endo-4DGS likely measure memorization of the sequence, not generalization to unseen lighting. The paper offers no experiment where the embedding is estimated from the test frame, or where held-out normal embeddings are used. This is not a minor issue: illumination correction for unseen surgical lighting is exactly the claimed deployment scenario.

Other soft spots are proportionate. All numbers come from a single run with no error bars. The comparison includes DarkGS and Gaussian-DK, which the paper itself marks as failing with NaN loss; that inflates the visual impression. The “maintaining geometric accuracy” claim is asserted but never directly evaluated—no Chamfer distance, F-score, or depth error on the corrected outputs. The EndoNeRF-EC dataset appears to be self-constructed with no public link, and the code repository is announced but not verifiable from the preprint.

The citation pattern is fine: the building blocks are all cited, including WildGaussians and ALeth-NeRF, and the baseline choices are reasonable. The thinking is clear and the engineering is honest; the problem is the protocol, not the intent.

Bottom line: this deserves a serious referee because the integration and the problem are both worthwhile, but the current evaluation does not support the headline claim. I would ask for multi-seed results, cleaned baselines (drop the NaN failures), a direct geometric metric, and a correction protocol that estimates embeddings from test inputs—or at least demonstrates that the training-frame embedding does not simply memorize the sequence.

Referee Report

3 major / 6 minor

Summary. Endo-4DGX extends 4D Gaussian Splatting for endoscopic scenes with variable illumination. It adds per-frame trainable illumination embeddings e ∈ R^{N×32}, routes frames into dark/bright region-aware enhancement networks that adjust Gaussian colors via an affine map, applies a spatial-aware quadratic adjustment, and uses a global exposure-control loss during training. The paper reports experiments on three datasets (EndoNeRF-EC, StereoMIS, C3VD), claiming large gains over reconstruction-plus-restoration baselines for illumination correction (e.g., 34.94 vs. 28.79 dB PSNR on Pulling) and improved uneven-illumination reconstruction, while running at 61 FPS with lower GPU memory. Ablations in Table 3 support each component, but the headline correction numbers depend on an evaluation protocol that uses a training-frame illumination embedding at test time.

Significance. The contribution is a practical, well-motivated extension of an existing dynamic-scene reconstruction backbone. The split dark/bright enhancement design and the exposure-control loss are sensible engineering choices, and the paper provides a code link, a detailed pipeline description, and deployment-relevant FPS/GPU-memory numbers that are often missing in this literature. If the illumination-correction evaluation can be made independent of training-frame embeddings and the geometry claim is backed by quantitative metrics, the method would be a useful step toward robust surgical scene reconstruction. As it stands, the experimental evidence is not yet sufficient to establish the central claims of illumination generalization and geometric accuracy; the main value is the proposed architecture rather than the currently demonstrated performance.

major comments (3)

[Section 3.2, Section 2.2, Eq. (3)-(4), Table 1] The illumination-correction protocol leaks the training embedding. Section 3.2 states that for EndoNeRF-EC illumination restoration "we utilize the first frame in the training data with normal-level illumination as the input embedding." However, in Section 2.2 the illumination embedding e is defined as a trainable per-frame parameter in R^{N×32}, updated jointly with the Gaussians under the color loss in Eq. (5). The corrected renderings for the test frames are therefore produced with an embedding that belongs to a training frame whose normal exposure was directly supervised, not with an embedding estimated from the test frame being corrected. On a temporally dense clip such as EndoNeRF-EC, this can amount to retrieving a memorized normal appearance from a nearby training frame. The Table 1 gains (e.g., 6.15 dB on Pulling and 10.76 dB on Cutting over EndoUIC+Endo-4DGS) consequently do not demonstrate correction of unseen lighting conditions. The authors should evaluate with (i) an embedding optimized for the test frame while the scene is frozen, (ii) held-out sequences that contain no normal-exposure training frames, and (iii) an analysis of sensitivity to the choice of reference normal frame.
[Abstract, Section 1, Eq. (6), Section 3] The claim that the method "maintains geometric accuracy" is unsupported by the reported experiments. Depth appears only in the training loss (Eq. (6)), and Tables 1-3 report PSNR, SSIM, LPIPS, FPS, and GPU memory, with no depth error, point-cloud accuracy, or other geometric metric. Since the datasets provide depth supervision (StereoMIS stereo matching, C3VD paired depth, and EndoNeRF rendering depth), the authors should report quantitative geometry results such as depth RMSE, absolute relative error, or Chamfer distance, or they should soften the geometry claim.
[Table 1, Section 3.3] Table 1 lists DarkGS and Gaussian-DK with a footnote marking them as failing with NaN loss during reconstruction, yet still reports their PSNR, SSIM, and LPIPS values. Including failed methods in the main comparison alongside successful baselines makes the reported margins harder to interpret; these methods should either be placed in a separate category or removed after explaining why they fail. In addition, all numbers in Tables 1-3 come from a single training run with no error bars. Given that EndoNeRF-EC consists of only two clips, the authors should report multiple seeds or otherwise quantify variance before claiming significant improvement over state-of-the-art methods.

minor comments (6)

[Table 3] The symbols "%" and "!" in Table 3 are not defined in the caption or the text, making the ablation rows difficult to interpret; please add an explicit legend.
[Eq. (3)] The notation "| β, γ= fregion(c, e)" in Eq. (3) is nonstandard and should be rewritten as "where β, γ = fregion(c, e)" for clarity.
[Table 1] In the row "CSEC+EndoGaussian" the citation is given as [11], but it should probably be [15] (EndoGaussian); please verify the reference.
[Table 2] The label "Pullling" in Table 2 is a typo and should read "Pulling."
[Section 2.2] The term "concealing network" appears to be a transfer of terminology from concealing-field NeRF methods; in this context it is confusing and should be replaced with a standard term such as "enhancement network."
[Contributions list] The phrase "exhausted experiments" should be "exhaustive experiments."

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method is trained with supervised photometric and depth losses against held-out ground-truth frames, and the training-frame embedding used for correction is an evaluation/protocol caveat rather than a construction-level reduction of the predicted output to its inputs.

full rationale

The paper's derivation chain is self-contained. Illumination embeddings e ∈ R^{N×32} are trainable per-frame parameters optimized jointly with the Gaussians (Section 2.2), and the color output is produced by explicit transformations: c_tone = β·c + γ with β,γ = f_region(c,e), then C̃_tone = C_tone + δ·C_tone(1−C_tone) with δ = f_spatial(e). These outputs are supervised by L_color against the ground-truth normal-exposure image Ĉ and by depth and exposure-control losses. No equation defines the predicted corrected image as the fitted embedding or as the fitted training-frame appearance; the rendered test-frame image is still a function of the deformed Gaussians at the test time and of the scene geometry. The only notable procedural issue is that illumination-restoration evaluation (Section 3.2) uses the first normal-level training frame's embedding as the input embedding, which may leak appearance information from training. However, this is an evaluation-protocol and generalization concern, not a circular one: the correction is not statistically forced to equal the ground-truth test image by construction, and the paper's central claim rests on held-out quantitative comparisons rather than on a self-referential definition. Self-citations to Endo-4DGS and EndoUIC are used as building blocks and baselines, not as an unverified uniqueness argument, and do not make the derivation circular.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The method is a learned system rather than a derivation, so its central claims rest on assumptions about the reliability of a pretrained illumination prior, the realism of a synthetically exposure-augmented endoscopic dataset, the accuracy of depth supervision, and the adequacy of a mean-brightness split into dark and bright frames. The free parameters are standard hyperparameters and a hand-chosen exposure target. No new physical entities are introduced.

free parameters (4)

Illumination embedding dimension k = 32
Hand-chosen dimension for the per-view illumination embeddings in Section 2.2.
Exposure level E = 0.6
Target average brightness in the global exposure control loss in Eq. 7.
Depth loss weight lambda_depth = 0.01
Weight chosen for the depth loss in Eq. 6.
TV loss weight lambda_tv = 0.01
Weight chosen for the total variation losses in Eqs. 5 and 6.

assumptions (4)

domain assumption The forward-reverse illumination estimation algorithm D from [29] produces a reliable lightness prior p = D(I) for classifying frames as bright or dark.
Used in Eq. 2 for the IC(I) split; no validation of this prior in endoscopic scenes is provided.
domain assumption The exposure-corrected EndoNeRF-EC dataset, created by simulating aperture changes with Adobe Camera Raw, approximates real endoscopic illumination variation.
The central correction experiments rely on this synthetic dataset; no independent dataset with paired normal-exposure ground truth is used.
domain assumption Depth supervision from stereo matching (OpenCV) on StereoMIS or a pretrained Depth Anything model is accurate enough for the depth loss.
The depth loss in Eq. 6 assumes these depth maps are reliable; inaccuracies could bias geometry, yet no geometric metrics are reported.
ad hoc to paper The mean-brightness threshold rule in Eq. 2 separates dark and bright frames sufficiently for the two-network design.
This heuristic is introduced for the paper's training imbalance solution and is not derived or independently validated.

invented entities (1)

Per-frame illumination embedding
purpose: A learned vector that conditions the region-aware and spatial-aware color correction on the brightness of each training view.
It is a learned representation rather than a physical entity, and its only observed support comes from the paper's own experiments. Appearance-style embeddings are common in radiance field literature, so this is a standard technique applied to illumination.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Endo-4DGX: Robust Endoscopic Scene Reconstruction and Illumination Correction with Gaussian Splatting." pith.science (2026). https://pith.science/paper/WXDQ5XW6

@misc{pith2026250623308,
  author       = {Pith},
  title        = {Pith review of: Endo-4DGX: Robust Endoscopic Scene Reconstruction and Illumination Correction with Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WXDQ5XW6}},
  note         = {Machine review of arXiv:2506.23308}
}

read the original abstract

Accurate reconstruction of soft tissue is crucial for advancing automation in image-guided robotic surgery. The recent 3D Gaussian Splatting (3DGS) techniques and their variants, 4DGS, achieve high-quality renderings of dynamic surgical scenes in real-time. However, 3D-GS-based methods still struggle in scenarios with varying illumination, such as low light and over-exposure. Training 3D-GS in such extreme light conditions leads to severe optimization problems and devastating rendering quality. To address these challenges, we present Endo-4DGX, a novel reconstruction method with illumination-adaptive Gaussian Splatting designed specifically for endoscopic scenes with uneven lighting. By incorporating illumination embeddings, our method effectively models view-dependent brightness variations. We introduce a region-aware enhancement module to model the sub-area lightness at the Gaussian level and a spatial-aware adjustment module to learn the view-consistent brightness adjustment. With the illumination adaptive design, Endo-4DGX achieves superior rendering performance under both low-light and over-exposure conditions while maintaining geometric accuracy. Additionally, we employ an exposure control loss to restore the appearance from adverse exposure to the normal level for illumination-adaptive optimization. Experimental results demonstrate that Endo-4DGX significantly outperforms combinations of state-of-the-art reconstruction and restoration methods in challenging lighting environments, underscoring its potential to advance robot-assisted surgical applications. Our code is available at https://github.com/lastbasket/Endo-4DGX.

Figures

Figures reproduced from arXiv: 2506.23308 by the authors.

**Figure 1.** Our method achieves state-of-the-art performance for robust reconstruction and illumination correction in the surgical scene with varying illumination. 1 Introduction Endoscopic procedures have emerged as a cornerstone of medical practice, particularly in the context of minimally invasive surgeries. In the recent development of robot-assisted minimally invasive surgeries, accurate reconstruction of the endoscopic … view at source ↗

**Figure 2.** Illustration of our proposed framework. We first preprocess the inputs for initialization, and then we train the Gaussians with our illumination embedding, regionaware enhancement, and spatial-aware adjustment for sub-area and global restoration. 2.2 Proposed Methodology: Endo-4DGX Illumination Embedding and Region-Aware Enhancement. We propose illumination embedding and region-aware enhancement to address the chal… view at source ↗

**Figure 3.** Qualitative Result on EndoNeRF-EC [21] Dataset. Results for DarkGS and Gaussian-DK are not presented due to the reconstruction failure. Our method provides the best reconstruction and illumination correction results for challenging illumination. C3VD [3] dataset is a video data collection designed for cardiovascular reconstruction. We utilize the cecum_t2_b and sigmoid_t2_a clips, including uneven illumination caus… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 21 canonical work pages

[1]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

Bai, L., Chen, T., Tan, Q., Nah, W.J., Li, Y., He, Z., Yuan, S., Chen, Z., Wu, J., Islam,M.,etal.:Endouic:Promptablediffusiontransformerforunifiedillumination correction in capsule endoscopy. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 296–306. Springer (2024)

work page 2024
[2]

In: International Conference on Medical Image Computing and Computer- Assisted Intervention

Bai, L., Chen, T., Wu, Y., Wang, A., Islam, M., Ren, H.: Llcaps: Learning to illu- minate low-light capsule endoscopy with curved wavelet attention and reverse dif- fusion. In: International Conference on Medical Image Computing and Computer- Assisted Intervention. pp. 34–44. Springer (2023)

work page 2023
[3]

Medical Image Analysis p

Bobrow, T.L., Golhar, M., Vijayan, R., Akshintala, V.S., Garcia, J.R., Durr, N.J.: Colonoscopy 3d video dataset with paired depth from 2d-3d registration. Medical Image Analysis p. 102956 (2023)

work page 2023
[4]

Bradski, G.: The opencv library. Dr. Dobb’s Journal of Software Tools (2000)

2000
[5]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

Chen, T., Lyu, Q., Bai, L., Guo, E., Gao, H., Yang, X., Ren, H., Zhou, L.: Lightdiff: Surgical endoscopic image low-light enhancement with t-diffusion. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 369–379. Springer (2024)

work page 2024
[6]

In: Proceedings of the AAAI Con- ference on Artificial Intelligence (2024)

Cui, Z., Gu, L., Sun, X., Ma, X., Qiao, Y., Harada, T.: Aleth-nerf: Illumination adaptive nerf with concealing field assumption. In: Proceedings of the AAAI Con- ference on Artificial Intelligence (2024)

work page 2024
[7]

In: 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI)

Garcia-Vega, A., Espinosa, R., Ramírez-Guzmán, L., Bazin, T., Falcón-Morales, L., Ochoa-Ruiz, G., Lamarque, D., Daul, C.: Multi-scale structural-aware exposure correction for endoscopic imaging. In: 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI). pp. 1–5. IEEE (2023)

work page 2023
[8]

CVPR (2020)

Guo, C., Li, C., Guo, J., Loy, C.C., Hou, J., Kwong, S., Runmin, C.: Zero-reference deep curve estimation for low-light image enhancement. CVPR (2020)

work page 2020

Show all 32 references

[9]

International Journal of Computer Assisted Radiology and Surgery pp

Hayoz, M., Hahne, C., Gallardo, M., Candinas, D., Kurmann, T., Allan, M., Sznit- man, R.: Learning how to robustly estimate camera pose in endoscopic videos. International Journal of Computer Assisted Radiology and Surgery pp. 1185—- 1192 (2023)

2023
[10]

arXiv preprint arXiv:2501.19319 (2025) 10 Y

Huang, Y., Cui, B., Bai, L., Chen, Z., Wu, J., Li, Z., Liu, H., Ren, H.: Advancing dense endoscopic reconstruction with gaussian splatting-driven surface normal- aware tracking and mapping. arXiv preprint arXiv:2501.19319 (2025) 10 Y. Huang et al

2025 arXiv
[11]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

Huang, Y., Cui, B., Bai, L., Guo, Z., Xu, M., Islam, M., Ren, H.: Endo-4dgs: Endo- scopic monocular scene reconstruction with 4d gaussian splatting. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 197–207. Springer (2024)

2024
[12]

ACM Transactions on Graphics42(4) (2023)

Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics42(4) (2023)

2023
[13]

NeurIPS (2024)

Kulhanek, J., Peng, S., Kukelova, Z., Pollefeys, M., Sattler, T.: WildGaussians: 3D gaussian splatting in the wild. NeurIPS (2024)

2024
[14]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Li, Y., Xu, K., Hancke, G.P., Lau, R.W.: Color shift estimation-and-correction for image enhancement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 25389–25398 (2024)

2024
[15]

arXiv preprint arXiv:2401.12561 (2024)

Liu, Y., Li, C., Yang, C., Yuan, Y.: Endogaussian: Gaussian splatting for de- formable surgical scene reconstruction. arXiv preprint arXiv:2401.12561 (2024)

2024 arXiv
[16]

In: Medical Image Computing and Computer Assisted Intervention–MICCAI 2020: 23rd International Conference, Lima, Peru, October 4–8, 2020, Proceedings, Part II 23

Ma, Y., Liu, Y., Cheng, J., Zheng, Y., Ghahremani, M., Chen, H., Liu, J., Zhao, Y.: Cycle structure and illumination constrained gan for medical image enhancement. In: Medical Image Computing and Computer Assisted Intervention–MICCAI 2020: 23rd International Conference, Lima, ...

2020
[17]

Commu- nications of the ACM65(1), 99–106 (2021)

Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Commu- nications of the ACM65(1), 99–106 (2021)

2021
[18]

Journal of Imaging10(1), 28 (2024)

Moghtaderi, S., Yaghoobian, O., Wahid, K.A., Lukong, K.E.: Endoscopic image enhancement: Wavelet transform and guided filter decomposition-based fusion ap- proach. Journal of Imaging10(1), 28 (2024)

2024
[19]

Mou, E.: Global and local enhancement of low-light endoscopic images-[v2] (2023)

2023
[20]

arXiv preprint arXiv:2304.07193 (2023)

Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)

2023 arXiv
[21]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

Wang, Y., Long, Y., Fan, S.H., Dou, Q.: Neural rendering for stereo 3d recon- struction of deformable tissues in robotic surgery. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 431–441. Springer (2022)

2022
[22]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 20310–20320 (June 2024)

2024
[23]

International Journal of Computer Vision131(4), 1022–1041 (2023)

Yang, K.F., Cheng, C., Zhao, S.X., Yan, H.M., Zhang, X.S., Li, Y.J.: Learning to adapt to light. International Journal of Computer Vision131(4), 1022–1041 (2023)

2023
[24]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Un- leashing the power of large-scale unlabeled data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10371–10381 (2024)

2024
[25]

In: Interna- tional Conference on Medical Image Computing and Computer-Assisted Interven- tion

Yang, S., Li, Q., Shen, D., Gong, B., Dou, Q., Jin, Y.: Deform3dgs: Flexible defor- mation for fast surgical scene reconstruction with gaussian splatting. In: Interna- tional Conference on Medical Image Computing and Computer-Assisted Interven- tion. pp. 132–142. Springer (2024)

2024
[26]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 20331– 20341 (2024) Endo-4DGX 11

2024
[27]

In: Computer Graphics Forum

Ye, S., Dong, Z.H., Hu, Y., Wen, Y.H., Liu, Y.J.: Gaussian in the dark: Real- time view synthesis from inconsistent dark images using gaussian splatting. In: Computer Graphics Forum. vol. 43, p. e15213. Wiley Online Library (2024)

2024
[28]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

Zha, R., Cheng, X., Li, H., Harandi, M., Ge, Z.: Endosurf: Neural surface re- construction of deformable tissues with stereo endoscope videos. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 13–23. Springer (2023)

2023
[29]

In: Computer graphics forum

Zhang, Q., Nie, Y., Zheng, W.S.: Dual illumination estimation for robust expo- sure correction. In: Computer graphics forum. vol. 38, pp. 243–252. Wiley Online Library (2019)

2019
[30]

Zhang, T., Huang, K., Zhi, W., Johnson-Roberson, M.: Darkgs: Learning neural illuminationand3dgaussiansrelightingforroboticexplorationinthedark.In:2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 12864–12871. IEEE (2024)

2024
[31]

International Journal of Computer Vision pp

Zuo, X., Samangouei, P., Zhou, Y., Di, Y., Li, M.: Fmgs: Foundation model em- bedded 3d gaussian splatting for holistic 3d scene understanding. International Journal of Computer Vision pp. 1–17 (2024)

2024
[32]

In: Proceedings of the 28th annual conference on Computer graphics and interactive techniques

Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: Surface splatting. In: Proceedings of the 28th annual conference on Computer graphics and interactive techniques. pp. 371–378 (2001)

2001

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

Bai, L., Chen, T., Tan, Q., Nah, W.J., Li, Y., He, Z., Yuan, S., Chen, Z., Wu, J., Islam,M.,etal.:Endouic:Promptablediffusiontransformerforunifiedillumination correction in capsule endoscopy. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 296–306. Springer (2024)

work page 2024

[2] [2]

In: International Conference on Medical Image Computing and Computer- Assisted Intervention

Bai, L., Chen, T., Wu, Y., Wang, A., Islam, M., Ren, H.: Llcaps: Learning to illu- minate low-light capsule endoscopy with curved wavelet attention and reverse dif- fusion. In: International Conference on Medical Image Computing and Computer- Assisted Intervention. pp. 34–44. Springer (2023)

work page 2023

[3] [3]

Medical Image Analysis p

Bobrow, T.L., Golhar, M., Vijayan, R., Akshintala, V.S., Garcia, J.R., Durr, N.J.: Colonoscopy 3d video dataset with paired depth from 2d-3d registration. Medical Image Analysis p. 102956 (2023)

work page 2023

[4] [4]

Bradski, G.: The opencv library. Dr. Dobb’s Journal of Software Tools (2000)

2000

[5] [5]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

Chen, T., Lyu, Q., Bai, L., Guo, E., Gao, H., Yang, X., Ren, H., Zhou, L.: Lightdiff: Surgical endoscopic image low-light enhancement with t-diffusion. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 369–379. Springer (2024)

work page 2024

[6] [6]

In: Proceedings of the AAAI Con- ference on Artificial Intelligence (2024)

Cui, Z., Gu, L., Sun, X., Ma, X., Qiao, Y., Harada, T.: Aleth-nerf: Illumination adaptive nerf with concealing field assumption. In: Proceedings of the AAAI Con- ference on Artificial Intelligence (2024)

work page 2024

[7] [7]

In: 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI)

Garcia-Vega, A., Espinosa, R., Ramírez-Guzmán, L., Bazin, T., Falcón-Morales, L., Ochoa-Ruiz, G., Lamarque, D., Daul, C.: Multi-scale structural-aware exposure correction for endoscopic imaging. In: 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI). pp. 1–5. IEEE (2023)

work page 2023

[8] [8]

CVPR (2020)

Guo, C., Li, C., Guo, J., Loy, C.C., Hou, J., Kwong, S., Runmin, C.: Zero-reference deep curve estimation for low-light image enhancement. CVPR (2020)

work page 2020