Towards Better Robustness: Pose-Free 3D Gaussian Splatting for Arbitrarily Long Videos

Nannan Li; Sheng Ye; Yong-jin Liu; Yu-Hui Wen; Zhen-Hui Dong

REVIEW 3 major objections 4 minor 7 references

Towards Better Robustness: Pose-Free 3D Gaussian Splatting for Arbitrarily Long Videos

T0 review · 3 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Rob-GS is a pose-free 3D Gaussian Splatting framework that reconstructs scenes from arbitrarily long, casually captured videos by tracking adjacent frames with a reprojection-flow loss and adaptively segmenting the video based on Gaussian…

desk verdict A genuine step forward for pose-free 3DGS on long videos, but the headline 'arbitrarily long' claim outruns the evidence: the evaluation protocol and the lack of any global anchor leave the strong version unverified. read the letter →

arxiv 2501.15096 v2 pith:EZSBQYA6 submitted 2025-01-25 cs.CV

Zhen-Hui Dong , Sheng Ye , Yu-Hui Wen , Nannan Li , Yong-Jin Liu This is my paper · ORCID

classification cs.CV

keywords novelviewsynthesis3DGaussianSplattingcameraposeestimationpose-freereconstructionmonoculardepthopticalflowarbitrarilylongvideoadaptivesegmentation

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

The paper introduces Rob-GS, a framework that reconstructs 3D scenes as Gaussian splats and estimates camera poses directly from arbitrarily long, casually captured video, skipping Structure-from-Motion entirely. The goal is to make pose-free novel view synthesis robust to the long sequences and erratic camera motion that defeat prior approaches. The paper argues that adjacent-frame tracking with a reprojection-flow loss, plus adaptive segmentation based on Gaussian visibility, together keep pose estimates stable and memory bounded. Because it works without known poses and scales to thousands of frames, the method targets practical deployment on phone-captured video.

What carries the argument

The central mechanism is adjacent pose tracking: a per-frame cloud is obtained by unprojecting the RGB image through an off-the-shelf monocular depth map, then fitted as 3D Gaussians. For the next frame, these Gaussians are frozen and the $\mathrm{SE}(3)$ camera pose is optimized so that the rendered image and a reprojected flow field match the observed image and RAFT optical flow. The paper derives analytic gradients of the color loss with respect to the pose on the $\mathrm{SE}(3)$ manifold and modifies the CUDA rasterizer to propagate them, making pose optimization fast. On top of that, a Gaussian visibility retention check—the fraction of a reference set of Gaussians that project into each new view—decides when to split the sequence into locally optimized segments, preventing memory overflow and global inconsistency.

What would settle it

Running Rob-GS on a long video through a textureless corridor with unreliable monocular depth, against motion-capture ground-truth poses, would refute the robustness claim if pose error grows with sequence length; a simpler test is to inject per-frame scale jitter into the depth maps and observe whether pose error spikes.

Watch

Extended reading notes

Core claim

Rob-GS claims that by fitting a set of Gaussians to a single frame, freezing them, and optimizing only the next camera pose under photometric and flow supervision, the relative motion between consecutive frames can be recovered reliably even when the camera moves fast or the trajectory is complex. A flow loss that matches reprojected depth-based flow to RAFT optical flow prevents pose tracking from failing in large-motion regions. To handle arbitrarily long input, the method monitors the fraction of previously visible Gaussians that remain in the current view; when that retention rate drops below a threshold, it closes the current segment, locally optimizes the Gaussians with densification, and starts the next segment with a five-frame overlap. The paper shows on Tanks and Temples, ScanNet, and a self-captured handheld-phone dataset that this yields higher rendering quality, lower pose error, and much shorter training time than Nope-NeRF, LocalRF, and CF-3DGS.

Load-bearing premise

The monocular depth maps from Depth Anything V2, aligned per frame only by a scale-invariant loss, must stay consistent across frames for unprojection and reprojection-flow tracking; if depth scale or shift drifts between frames, pose and geometry drift because the system has no metric anchor or global bundle adjustment.

Editorial extensions

If this is right

SfM pre-processing can be dropped entirely for long-video reconstruction, removing a frequent failure point on textureless or low-overlap footage.
Memory usage stays bounded by segment size, so arbitrarily long sequences can be processed on a single GPU without out-of-memory failure.
Casually captured handheld footage with fast, erratic motion becomes tractable for pose-free novel view synthesis, expanding the range of deployable scenes.
Training time drops to roughly one quarter of LocalRF's on the tested benchmarks, making high-quality reconstruction of long sequences more practical.

Reading between the lines

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

The paper leaves loop closure to future work; if a video revisits a region, the per-segment optimization has no mechanism to correct drift when the loop closes, so global consistency over very long traversals may still degrade without an added global alignment stage.
Because pose tracking relies on consistent per-frame monocular depth, adding a multi-view consistency check on depth could reduce the method's dependence on the single-frame prior and improve robustness further.
The visibility retention threshold is a cheap keyframe-selection signal that could be transferred to other incremental reconstruction systems, such as SLAM or large-scale NeRF pipelines, as a trigger for local optimization.

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

Two things to know. First, this is a real engineering contribution to pose-free 3DGS on long sequences, not a repackaging of CF-3DGS. Second, the evaluation protocol undercuts the headline numbers: test-view poses are fitted against the frozen model using the test image itself, so the PSNR/LPIPS values are reconstruction-under-fitted-pose, not true novel-view synthesis. That doesn't kill the paper, but you should read Tables 1-3 with that caveat.

What's new: the adjacent pose tracking with a reprojection-flow loss (RAFT optical flow supervising unprojected/projected correspondences) and the Gaussian visibility-retention segmentation heuristic. The camera Jacobian extension into the CUDA rasterizer is non-trivial and well-derived. The ablations support both components: removing the segmentation on Tanks and Temples drops PSNR from 29.05 to 21.12, and removing the flow loss produces visible ghosting. That is convincing evidence the method itself works as described.

Where it's soft: the 'arbitrarily long' claim. The system has no loop closure (they say so in Sec. 5), and monocular depth is aligned only per-frame via scale/shift in Eq. 5. The flow loss couples translation scale to unprojected depth scale, so pose drift over thousands of frames is not bounded by anything in the method. The experiments top out around 1000-2000 frames, and pose evaluation uses COLMAP pseudo-ground-truth on Tanks and Temples, which they themselves acknowledge is inaccurate for long sequences. So the strong version of the headline is not demonstrated. There is also no code or data release, no error bars, and the self-captured dataset is not public. The test-pose fitting is the biggest methodological caveat; I'd want a comparison with held-out views without pose fitting, or at least a discussion of how much the fitted poses inflate the numbers.

Minor: the per-frame depth alignment is a known limitation in monocular-depth pose methods, and it directly qualifies the 'robust' claim. Not fatal, but worth flagging.

Bottom line: a solid, well-written paper for the pose-free 3DGS subfield. A serious referee should engage with the evaluation protocol and the drift argument. I'd send it to review with those issues clearly requested, and I'd cite the flow-loss and segmentation ideas in my own work.

Referee Report

3 major / 4 minor

Summary. Rob-GS is a pose-free 3D Gaussian Splatting framework for long video sequences. It progressively estimates camera poses by fitting per-frame Gaussians from monocular depth, then tracking each new frame with a color loss and a flow loss that compares reprojection flow to RAFT optical flow. A visibility-retention check splits the video into overlapping segments and a local 3DGS is optimized per segment. Experiments on Tanks and Temples, ScanNet, and a self-captured phone dataset compare against Nope-NeRF, LocalRF, and CF-3DGS, reporting higher PSNR/SSIM/LPIPS and lower training time, plus RPE pose errors against COLMAP. Ablations show that the flow loss and the segment-wise optimization scheme improve quality.

Significance. If the central robustness and evaluation concerns were resolved, Rob-GS would represent a practical step toward SfM-free 3DGS on long, casually captured videos: the flow-induced adjacent tracking and adaptive segmentation are reasonable design choices, the CUDA-level pose Jacobians are a useful engineering contribution, and the benchmarks cover substantially longer sequences than prior pose-free 3DGS work. However, the pose-fitted test-view evaluation and the lack of any drift or loop-closure analysis mean that the main claims, as currently supported, are weaker than the abstract's 'arbitrarily long videos' and 'robust pose estimation' suggest. The paper's own limitations section acknowledges the absence of loop closure, which is exactly the mechanism needed to bound long-sequence drift.

major comments (3)

[4.1.2 (Tables 1-3)] Section 4.1.2 states that for test views the trained 3DGS is frozen and the camera pose is optimized directly against each test image, and that metrics are computed under these optimized poses. As a result, the reported PSNR, SSIM, and LPIPS values in Tables 1-3 measure how well the model can be made to reproduce each test image after fitting its pose, not true novel-view synthesis from unseen poses. Because pose fitting can compensate for geometric or photometric errors in the reconstructed scene, the headline claim of superior novel-view synthesis over Nope-NeRF, LocalRF, and CF-3DGS is not supported by the current evaluation protocol. Please report results with poses fixed by COLMAP or ground truth, or clearly re-label these as pose-fitted reconstruction metrics.
[3.3.1-3.3.2, Eqs. (5)-(8), and Section 5] The adjacent pose tracker aligns each monocular depth map only by a per-frame scale-invariant loss (Eq. 5) and couples the estimated translation to the per-frame depth scale through the unprojection in Eqs. (6)-(7) and the flow loss in Eq. (8). Since Section 5 states that no loop closure is incorporated and the paper does not perform global bundle adjustment, any drift in the monocular depth scale or shift can be absorbed into the translation estimate and accumulate over arbitrarily long sequences. The 1000-2000 frame experiments do not measure absolute trajectory error, so the central claim of robust pose estimation for arbitrarily long videos is not established. Please add a drift analysis on sequences with ground-truth poses, or add a loop-closure/global alignment mechanism, to support the claim.
[4.2, Table 4] The pose evaluation uses COLMAP poses as pseudo-ground-truth and reports RPE only on Tanks and Temples. The paper itself notes that COLMAP is not accurate for long video sequences, and no estimate of COLMAP's own drift or error bars is provided; moreover, no pose metrics are reported for ScanNet or the self-captured dataset. Thus the claim that Rob-GS produces robust pose estimates on the harder, longer sequences is not quantitatively supported. Please evaluate against an independent pose source, or at least quantify the accuracy of the pseudo-ground-truth reference.

minor comments (4)

[3.1] There is a typo in Section 3.1: 'adaptive sementation strategy' should be 'adaptive segmentation strategy'.
[3.3.1-3.3.2] The symbol D_i is used both for the monocular depth map in Eq. (5) and for the depth map used in the unprojection of Eq. (6); please use distinct notation, such as D_i^mono and D_i, to avoid ambiguity.
[Figure 3] Figure 3 shows example values R_c = 0.1 and R_c = 0.5, but the text defines R_c as a retention rate; please clarify whether these are thresholds used in the experiments and specify how the threshold was selected.
[4.2, Table 4] The RPE metric is cited but not defined; please state the exact formulas for RPE_t and RPE_r so that the reported numbers are interpretable.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Rob-GS's pose tracking and segmentation are derived from external monocular depth and optical-flow priors with standard SE(3) calculus; the only same-author citation is descriptive and non-load-bearing.

full rationale

The claimed derivation chain is self-contained. Single-frame Gaussian fitting (Eqs. 4-5) uses the external Depth Anything V2 prior as input; adjacent-pose tracking (Eqs. 6-9) is a conventional unproject-project flow constraint supervised by RAFT optical flow and photometric consistency, and no target quantity is defined in terms of the output pose or render. The camera Jacobians (Eqs. 10-12) are standard Lie-algebra derivatives cited to external sources (Deray and Solà 2020; Matsuki et al. 2024), not to the authors' own work. The only same-author reference (Ye et al. 2024) appears in Related Works as one example of 3DGS extensions and plays no role in the method or ablation. The admitted absence of loop closure (Sec. 5) and the test-time pose optimization for evaluation (Sec. 4.1.2) are genuine empirical and robustness caveats, but they are not circular derivation: the method's equations and inputs would be unchanged even if those caveats were addressed. Therefore no step of the derivation reduces by construction to its own inputs.

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

The central claim rests on off-the-shelf monocular depth and optical flow models plus several hand-set hyperparameters. The method introduces no new physical entities. The main unstated premise is that camera intrinsics are known, which is not discussed. The scale ambiguity of monocular depth is handled only per image by a scale-invariant loss, leaving cross-frame scale drift as an open risk.

free parameters (5)

depth loss weight lambda_d
Used in Eq. 4 and Eq. 13 for single-image fitting and local optimization; exact value not reported in the main text.
flow loss weight lambda_f
Used in Eq. 9 for pose tracking; exact value not reported in the main text.
segmentation threshold R_c
The visibility retention rate threshold triggers new segments (Sec. 3.4.1); Fig. 3 shows examples of 0.1 and 0.5 but the exact threshold and its sensitivity are not given.
segment overlap length = 5 frames
A fixed overlap between segments is used to ensure seamless transitions (Sec. 3.4.2), which affects final quality.
densification stop ratio = 0.7
Densification stops at 70% of total local optimization iterations (Sec. 4.1.3), a hand-set schedule that affects memory and quality.

assumptions (5)

domain assumption Monocular depth from Depth Anything V2 provides a reliable and sufficiently consistent depth prior for every frame.
Used in Eq. 4-5 for single-frame Gaussian fitting and in Eq. 6-7 for unprojection in the reprojection flow. If depth scale or shift drifts across frames, pose tracking and reconstruction inherit the error.
domain assumption RAFT optical flow gives correct dense correspondences between adjacent frames even under large or fast camera motion.
The flow loss in Eq. 8 supervises reprojection flow; incorrect flow would pull the pose estimate toward the wrong alignment.
domain assumption Camera intrinsics are known or fixed.
Projection and unprojection in Eq. 2 and Eqs. 6-7 require focal length and principal point, but the paper never states how intrinsics are obtained in the pose-free setting.
domain assumption Adjacent video frames have significant overlap and small inter-frame pose changes.
The adjacent pose tracking design and identity initialization of the first pose (Sec. 3.3) rely on temporal continuity; low-overlap footage would break tracking.
domain assumption A five-frame overlap between adjacent segments is sufficient to prevent visible seams.
Sec. 3.4.2 uses a fixed overlap to blend segments; if scene content changes faster than the overlap, artifacts may appear.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Towards Better Robustness: Pose-Free 3D Gaussian Splatting for Arbitrarily Long Videos." pith.science (2026). https://pith.science/paper/EZSBQYA6

@misc{pith2026250115096,
  author       = {Pith},
  title        = {Pith review of: Towards Better Robustness: Pose-Free 3D Gaussian Splatting for Arbitrarily Long Videos},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EZSBQYA6}},
  note         = {Machine review of arXiv:2501.15096}
}

read the original abstract

3D Gaussian Splatting (3DGS) has emerged as a powerful representation due to its efficiency and high-fidelity rendering. 3DGS training requires a known camera pose for each input view, typically obtained by Structure-from-Motion (SfM) pipelines. Pioneering works have attempted to relax this restriction but still face difficulties when handling long sequences with complex camera trajectories. In this paper, we propose Rob-GS, a robust framework to progressively estimate camera poses and optimize 3DGS for arbitrarily long video inputs. In particular, by leveraging the inherent continuity of videos, we design an adjacent pose tracking method to ensure stable pose estimation between consecutive frames. To handle arbitrarily long inputs, we propose a Gaussian visibility retention check strategy to adaptively split the video sequence into several segments and optimize them separately. Extensive experiments on Tanks and Temples, ScanNet, and a self-captured dataset show that Rob-GS outperforms the state-of-the-arts.

Figures

Figures reproduced from arXiv: 2501.15096 by the authors.

**Figure 1.** Camera pose estimations and novel view synthesis. We propose Rob-GS to reconstruct scenes from arbitrarily long, casually captured videos without known camera poses. Rob-GS enables robust pose estimations (first row) and high-quality renderings (second row). 3D Gaussian Splatting (3DGS) has emerged as a powerful representation due to its efficiency and high-fidelity rendering. 3DGS training requires a known camera p… view at source ↗

**Figure 2.** The overall framework of Rob-GS. We use 3D Gaussians (Sec. 3.2) as the scene representation, and progressively estimate camera poses using a robust tracking approach (Sec. 3.3) that leverages adjacent image pairs. To handle long video sequence, we design an adaptive segmentation scheme (Sec. 3.4) to split the video sequence into several local segments and optimize them individually. them separately. These designs en… view at source ↗

**Figure 3.** Gaussian visibility retention check strategy. The blue camera exhibits a higher visibility retention rate that exceeds the threshold. Therefore, the blue camera’s frame should be added to the current segment, while the red camera’s frame should be assigned to a new segment. we further derive the following equations: 𝜕𝜇′ 𝜕𝑇𝑊𝐶 = 𝜕𝜇′ 𝜕𝜇 D𝜇 D𝑇𝑊𝐶 , 𝜕Σ ′ 𝜕𝑇𝑊𝐶 = 𝜕Σ ′ 𝜕 𝐽 𝜕 𝐽 𝜕𝜇 D𝜇 D𝑇𝑊𝐶 + 𝜕Σ ′ 𝜕𝑊 D𝑊 D𝑇𝑊𝐶 , (11) D𝜇 D𝑇𝑊𝐶 = … view at source ↗

Figures from the paper (6 more)

**Figure 4.** Figure 4: Qualitative comparison for novel view synthesis on Tanks and Temples. Our approach produces sharper details than other baselines. 𝐼𝑖 , using its associated monocular depth map 𝐷𝑖 and camera pose 𝑇𝑖 . For scenes with a large spatial extent, we additionally project point…

**Figure 5.** Figure 5: Qualitative comparison for novel view synthesis on self-captured dataset. Our approach produces the most realistic rendering results among other baselines. Highly recommend zooming in for better comparison. methods. We visualize the camera trajectories of some scenes f…

**Figure 6.** Figure 6: The visualizations of our ablation study. Artifacts will appear when any of our proposed components are removed. Our Rob-GS achieves superior rendering quality and faster training speed across all datasets. Compared to LocalRF, our approach shows significant improveme…

**Figure 7.** Figure 7: Qualitative results of novel view synthesis on ScanNet dataset. Our Rob-GS is able to preserve fine-grained datails. Nope-NeRF CF-3DGS LocalRF Ours Ground-Truth [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

**Figure 8.** Figure 8: Qualitative results of novel view synthesis on Tanks and Temples and self-captured datasets. Zoom in for better comparison. Gate Pavillion2 Horse Barn [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

**Figure 9.** Figure 9: Camera trajectories estimated by Rob-GS. The first two scenes are from self-captured dataset, the latter two from Tanks and Temples [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

7 extracted references · 6 canonical work pages

[4]

In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XVI , Vol

Citygaussian: Real-time high-quality large-scale scene rendering with gaus- sians. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XVI , Vol. 15074. Springer, Heidelberg, Germany, 265–282. Hidenobu Matsuki, Riku Murai, Paul H. J. Kelly, and Andrew J. Davison. 2024. Gaussian Splatting...

work page 2019
[7]

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

ICON: Incremental CONfidence for Joint Pose and Radiance Field Optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, US, 5406–5417. Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. 2021. NeRF–: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07...

arXiv 2021
[2001]

InProceedings of the 28th annual conference on Computer graphics and interactive techniques

Unstructured lumigraph rendering. InProceedings of the 28th annual conference on Computer graphics and interactive techniques . ACM, US, 425–432. Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. 2022. TensoRF: Tensorial Radiance Fields. InComputer Vision - ECCV 2022 - 17th European Conference, Tel A viv, Israel, October 23-27, 2022, Proceedi...

work page 2020
[2012]

In 2012 IEEE/RSJ international conference on intelligent robots and systems

A benchmark for the evaluation of RGB-D SLAM systems. In 2012 IEEE/RSJ international conference on intelligent robots and systems . IEEE, US, 573–580. Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. 2022. Block-nerf: Scalable large scene neural view synthesis. In ...

work page 2012
[2020]

IEEE transactions on pattern analysis and machine intelligence 44, 3 (2020), 1623–1637

Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence 44, 3 (2020), 1623–1637. Johannes L Schönberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. 2016. Pixelwise view selection for unstructured multi-view stereo. In Computer Vision– ECCV 20...

work page 2020
[2023]

In Proceedings of the IEEE/CVF International Conference on Computer Vision

Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision . IEEE, US, 19697–19705. Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, and Victor Adrian Prisacariu. 2023. Nope-nerf: Optimising neural radiance field with no pose prior. In Proceedings of the IEEE/CVF Conference on Com...

work page 2023
[2024]

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

Free-SurGS: SfM-Free 3D Gaussian Splatting for Surgical Scene Reconstruction. In International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, Heidelberg, Germany, 350–360. Yoonwoo Jeong, Seokjun Ahn, Christopher Choy, Anima Anandkumar, Minsu Cho, and Jaesik Park. 2021. Self-calibrating neural radiance fields. In Procee...

work page 2023

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [4]

In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XVI , Vol

Citygaussian: Real-time high-quality large-scale scene rendering with gaus- sians. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XVI , Vol. 15074. Springer, Heidelberg, Germany, 265–282. Hidenobu Matsuki, Riku Murai, Paul H. J. Kelly, and Andrew J. Davison. 2024. Gaussian Splatting...

work page 2019

[2] [7]

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

ICON: Incremental CONfidence for Joint Pose and Radiance Field Optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, US, 5406–5417. Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. 2021. NeRF–: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07...

arXiv 2021

[3] [2001]

InProceedings of the 28th annual conference on Computer graphics and interactive techniques

Unstructured lumigraph rendering. InProceedings of the 28th annual conference on Computer graphics and interactive techniques . ACM, US, 425–432. Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. 2022. TensoRF: Tensorial Radiance Fields. InComputer Vision - ECCV 2022 - 17th European Conference, Tel A viv, Israel, October 23-27, 2022, Proceedi...

work page 2020

[4] [2012]

In 2012 IEEE/RSJ international conference on intelligent robots and systems

A benchmark for the evaluation of RGB-D SLAM systems. In 2012 IEEE/RSJ international conference on intelligent robots and systems . IEEE, US, 573–580. Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. 2022. Block-nerf: Scalable large scene neural view synthesis. In ...

work page 2012

[5] [2020]

IEEE transactions on pattern analysis and machine intelligence 44, 3 (2020), 1623–1637

Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence 44, 3 (2020), 1623–1637. Johannes L Schönberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. 2016. Pixelwise view selection for unstructured multi-view stereo. In Computer Vision– ECCV 20...

work page 2020

[6] [2023]

In Proceedings of the IEEE/CVF International Conference on Computer Vision

Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision . IEEE, US, 19697–19705. Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, and Victor Adrian Prisacariu. 2023. Nope-nerf: Optimising neural radiance field with no pose prior. In Proceedings of the IEEE/CVF Conference on Com...

work page 2023

[7] [2024]

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

Free-SurGS: SfM-Free 3D Gaussian Splatting for Surgical Scene Reconstruction. In International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, Heidelberg, Germany, 350–360. Yoonwoo Jeong, Seokjun Ahn, Christopher Choy, Anima Anandkumar, Minsu Cho, and Jaesik Park. 2021. Self-calibrating neural radiance fields. In Procee...

work page 2023