acceptodds
Withdrawn from ICLR 2027

FVVSplat: Feed-Forward Streamable Free-Viewpoint Video Reconstruction from Uncalibrated Inputs

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Streamable Free-Viewpoint Video (FVV) reconstruction aims to recover dynamic 3D scenes from multi-view observations, requiring high-quality, on-the-fly reconstruction under strict constraints. Existing methods rely on known camera parameters and slow per-scene optimization, which hinders practical application. In this work, we introduce FVVSplat, a Feed-Forward Streamable Free-Viewpoint Video Reconstruction framework that instantly predicts dynamic 3D Gaussian Splatting representations from uncalibrated multi-view video streams. There are three key innovations are proposed. 1) an adaptive radius sampling mechanism for accurate Gaussian prediction from uncalibrated inputs; 2) a bidirectional deformation that ensures reliable, persistent Gaussians across frames and mitigates long-term error accumulation; 3) an efficient attention is integrated into both static prediction and dynamic deformation with linear complexity, substantially reducing computational cost while maintaining reconstruction quality across video resolutions. Extensive experiments on multiply streamable FVV reconstruction benchmarks demonstrate that FVVSplat significantly improves both reconstruction quality and inference time, with a 15x speedup and 2.16 dB PSNR gain over existing optimization-based methods on the N3DV dataset. Our source code and models will be available at http://github.com.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Feed-forward Reconstruction, Free-viewpoint Video, Streamable, Uncalibration

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_Evj6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FVVSplat aims to reconstruct dynamic Gaussian scenes from uncalibrated multi-view video streams using adaptive position sampling, bidirectional deformation, and linear attention. The reported quality and speed improvements are promising, and the problem has clear practical value. However, the distinction from StreamSplat is insufficiently established, several essential parts of the multi-view pipeline are unclear, and the experiments do not fully support the claims about streaming stability and efficiency.

Critical strengths
  1. The paper addresses a useful problem. Reducing camera calibration requirements and per-frame optimization would make free-viewpoint video reconstruction easier to deploy.

  2. The reported results are strong across three benchmarks. On N3DV, the method reaches 34.82 dB PSNR with a reported processing time of 0.26 seconds per frame.

  3. The component ablations are useful. Table 3 examines all three proposed designs, and the visual comparisons show improvements in several challenging regions.

Critical weaknesses
  1. The novelty needs stronger justification. StreamSplat already includes probabilistic position prediction, bidirectional deformation, and online Gaussian fusion. Equation 9 also follows established linear attention. The paper needs clearer attribution and direct feed-forward comparisons. The contribution list even retains the name “StreamSplat.”

  2. The uncalibrated multi-view pipeline is underspecified. Section 4 mainly describes individual frames and temporal pairs. It does not clearly explain how different cameras are aligned, how their Gaussians share coordinates, or how target viewpoints are specified under the orthographic formulation.

  3. The sampling bounds appear inconsistent with the depth mapping. The paper uses g(z)=2/(1+z)g(z)=2/(1+z), but allows a sampling radius up to 1.2. The stated sampling interval can therefore cross z=−1z=-1, producing a singularity or negative depth. Any additional constraint needs to be specified.

  4. The evaluation split is not sufficiently clear. Section 5.1 states that official splits are used, but does not specify input camera counts, held-out target views, or whether scenes overlap between training and testing. These details are essential for interpreting the large gains.

  5. Long-term streaming stability is not demonstrated. Framewise image metrics and selected visual examples do not establish reduced temporal drift. The paper also lacks a clear pruning policy and measurements of Gaussian count or memory usage as the sequence grows.

  6. The efficiency discussion overstates the evidence. In Table 4, efficient attention is slower at 480P, 720P, and 960P; it becomes faster only at 1080P. The timing scope is also unclear, including whether depth estimation, DINOv2 features, and rasterization are all counted.

Questions for authors
  1. Which components are new relative to StreamSplat, and how does FVVSplat compare under matched conditions?

  2. How are camera coordinates aligned and held-out target viewpoints rendered without input calibration?

  3. What prevents sampled depth offsets from crossing the singularity at z=−1z=-1?

  4. Please specify the camera, scene, and temporal splits, including whether target images enter the model.

  5. Can you provide long-sequence temporal results and a complete runtime and memory breakdown?

Reviewer_mudzOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

FVVSplat proposes feed-forward reconstruction of dynamic 3D Gaussians from uncalibrated multi-view video streams, using adaptive-radius position sampling, bidirectional deformation, and linear attention. It reports improved reconstruction quality and processing time over optimization-based methods on N3DV, MeetRoom, and VRU.

Critical strengths

N/A, see weaknesses

Critical weaknesses

This paper appears to be a lightly modified rewrite of StreamSplat [1], with extensive textual and structural reuse that is not adequately acknowledged. [1] Wu, Z., Yan, Q., Yi, X., Wang, L., & Liao, R. (2026, April). Streamsplat: Towards online dynamic 3d reconstruction from uncalibrated video streams. In International Conference on Learning Representations (Vol. 2026, pp. 123114-123137).

Five concrete examples:

  1. Introduction and contribution bullets (p. 2, lines 73–95). The first contribution still states “We introduce StreamSplat” (line 85). The preceding design narrative closely follows StreamSplat's introduction, including its explanations of geometric uncertainty and bidirectional deformation.

  2. Related work (Sections 2.1–2.2). Lines 101–115 closely reproduce StreamSplat Section 2's discussion. More decisively, lines 138–139 repeat its closing sentence with only “StreamSplat” replaced by “FVVSplat.”

  3. Architecture figure (Figure 2). This closely adapts StreamSplat Figure 9's distinctive layout, DINOv2 path, paired tokens, decoder, upsampler, and two heads, with substituted examples and attention labels.

  4. Method descriptions (Sections 4.1–4.2). The canonical-space paragraph (lines 208–211) and dynamic-decoding paragraph (255–259) closely track StreamSplat Sections 3.1–3.2 in wording and order.

  5. Training text and equations (Section 4.4). Equations (10)–(12) reproduce StreamSplat Eqs. (2)–(4), respectively, alongside near-verbatim stage-one prose (lines 299–305). The evidence is the matching explanations and ordered formulation, beyond merely sharing common loss terms.

Although StreamSplat is cited, the manuscript does not identify the adapted figure or repeated passages, and presents overlapping components as its own contributions. This pattern raises a serious research-integrity concern and warrants examination by the AC.

Questions for authors

See weaknesses.

Reviewer_CUPDOpenReview ↗
Rating
1
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

FVVSplat proposes a feed-forward framework for streaming free-viewpoint video reconstruction from synchronized, uncalibrated multi-view inputs. It predicts Gaussian representations with a static encoder, estimates bidirectional deformation between consecutive observations, and combines the resulting Gaussians through opacity modulation. The paper presents adaptive-radius position sampling and efficient attention as mechanisms for improving reconstruction quality and computational efficiency. Experiments on N3DV, MeetRoom, and VRU report substantial quality and processing-time improvements over optimization-based streaming methods.

Critical strengths
  1. The paper addresses a practically relevant reconstruction setting. Reconstructing dynamic scenes incrementally from uncalibrated multi-view videos, without per-scene optimization, is relevant to interactive free-viewpoint applications and reduces reliance on camera calibration.

  2. The reported quality and processing-time improvements are substantial. The paper reports PSNR gains of 2.16 dB, 2.68 dB, and 4.95 dB over the strongest reported PSNR baselines on N3DV, MeetRoom, and VRU, respectively, with processing times of 0.17–0.28 seconds per frame.

Critical weaknesses

1. Serious concerns about possible textual plagiarism from StreamSplat.

The manuscript contains multiple passages that are identical or closely parallel to StreamSplat: Towards Online Dynamic 3D Reconstruction from Uncalibrated Video Streams, arXiv v2. The overlap extends across the method overview, related work, architecture descriptions, fusion explanation, and training procedure.

FVVSplat locationStreamSplat counterpartObserved correspondence
p. 6, lines 299–301§3.3, Stage 1 trainingThree consecutive explanatory sentences have identical wording.
p. 4, lines 200–202§3, method overviewThe method-summary sentence is identical apart from section references.
p. 3, lines 139–140§2, closing contrast sentenceThe sentence is identical except for replacing the model name.
p. 4, lines 208–212§3.1, canonical spaceClosely matching wording and rationale, with limited edits.
pp. 4–5, lines 213–220§3.1, static encodingClosely matching encoding description and identical position/depth mapping.
p. 5, lines 255–259§3.2, deformation decodingClosely parallel paragraph with changes to attention, wording, and notation.
p. 6, lines 273–275§3.2, Gaussian fusionSubstantial consecutive phrase overlap in the fusion explanation.
p. 7, lines 360–363§4.1, implementation detailsMatching learning rates, warm-ups, weight decay, and gradient clipping.

The strongest examples are the three consecutive identical sentences describing Stage 1, the method overview with only section references changed, and the related-work sentence with only the model name replaced. Additionally, FVVSplat’s contribution statement on p. 2, line 085 still begins “We introduce StreamSplat”. Although this wording error alone would not establish plagiarism, it reinforces the need to investigate the broader textual overlap.

2. The provenance and attribution of the shared methodology are insufficiently explained.

The shared design includes orthographic canonical space, pixel-aligned Gaussian encoding, probabilistic position sampling, bidirectional deformation, DINOv2-conditioned decoding, opacity-based fusion, and static-then-dynamic training. Corresponding reconstruction and depth-loss formulations also appear in both papers. StreamSplat, §3

FVVSplat cites StreamSplat in related work and the projection formulation. My concern is therefore not a complete absence of citation, but the failure to clearly distinguish inherited components from original contributions alongside extensive textual overlap. In particular, bidirectional deformation is presented as a key proposed design without adequately explaining its relationship to the existing method.

Adaptive-radius sampling, efficient attention, the multi-view setting, and the modified opacity formulation are identifiable differences. However, these changes do not explain the duplicated passages or resolve the attribution of the surrounding framework. Taken together, the textual correspondences and unclear methodological provenance constitute a serious concern about possible plagiarism and misattribution that requires a direct explanation.

Questions for authors

Questions for Authors

  1. Please directly address the apparent textual reuse from StreamSplat. How do you explain the three identical Stage 1 sentences, the method overview with only section references changed, and the related-work sentence with only the model name replaced? Please explain the origin of these passages and how their reuse was attributed.

  2. Please clarify the full relationship between FVVSplat and StreamSplat. Which architectural components, formulations, training procedures, code, or pretrained weights, if any, were reused or adapted? Please distinguish inherited components from new contributions explicitly.

  3. Why does the contribution statement on p. 2, line 085 introduce StreamSplat rather than FVVSplat? Please explain this alongside the other textual correspondences.

  4. What is specifically new in the bidirectional deformation and Gaussian fusion components relative to StreamSplat, and why are these presented as proposed contributions without a clear account of that relationship?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.