acceptodds
Under review as a conference paper at ICLR 2027

Motion from Blur: Intra-Exposure Trajectory Recovery for Feed-Forward 3D Deblurring

Abstract

Recent feed-forward models estimate camera poses and scene geometry from multi-view images in a single forward pass, enabling fast 3D reconstruction. However, their performance degrades markedly on motion-blurred inputs. We find that their pose estimates exhibit a systematic bias aligned with the motion-blur direction, which persists even after fine-tuning on blurred data, revealing a mismatch between single-pose prediction and trajectory-based blur formation. We therefore propose MoB3R, a feed-forward framework that recovers the intra-exposure camera trajectory and uses it to reconstruct sharp 3D Gaussians. Rather than treating blur only as an artifact, MoB3R treats its spatial pattern as a signal of camera motion: the geometry stage recovers exposure trajectories with additional motion tokens, and the appearance stage derives a per-pixel motion field to guide sharp reconstruction through cross-attention, supervised by a cycle reblur loss. Experiments on synthetic and real blurred scenes show improved pose estimation and novel-view rendering across blur levels, while reconstructing in seconds rather than the hours required by per-scene optimization.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
3DGS, 3D Reconstruction, Low-level vision, Image Deblur

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_nC8pOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

MoB3R reconstructs sharp 3D Gaussians from sparse, unposed, motion-blurred images. It extends WorldMirror with exposure-trajectory predictions and reversal-invariant supervision, then uses the recovered motion to guide appearance reconstruction and reblur consistency. Experiments demonstrate improved pose estimation and rendering on synthetic and real blur, with optional per-scene refinement providing further gains.

Critical strengths
  1. The method connects exposure-motion estimation and sharp reconstruction in a coherent pipeline. Table 5 supports the contributions of reversal-invariant supervision, motion guidance, and reblur consistency beyond simply fine-tuning the backbone.
  2. Improvements are consistent across view counts and both synthetic and real data. At 12 views, novel-view PSNR exceeds fine-tuned WorldMirror by 1.40 dB on DL3DV and 1.72 dB on ExBluRF, supporting practical value for sparse reconstruction.
Critical weaknesses

The evidence for accurate exposure-motion recovery is less conclusive than the rendering results. Camera-label construction for interpolated subframes is unspecified, and Table 9 evaluates globally aligned poses without isolating motion relative to the midpoint. Its gains may therefore partly reflect improved midpoint estimation. The stated reason for omitting real trajectory evaluation also needs clarification: Appendix C.4 says ExBluRF lacks trajectory references, whereas the original ExBluRF paper describes and evaluates such references.

Questions for authors
  1. How are camera labels assigned to interpolated exposure subframes? Under a reversal-invariant relative-motion metric, how do the predicted trajectories compare with simply repeating the predicted midpoint pose?
  2. Which ExBluRF release or subset was used, and why are the trajectory references described in the original paper unavailable for this evaluation?
Reviewer_bcvLOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes MoB3R, a feed-forward framework that recovers intra-exposure camera trajectories from blurred multi-view images and uses motion guidance to reconstruct sharp 3D Gaussians. Experiments on synthetic and real-world blur demonstrate improved pose estimation and novel-view synthesis over the evaluated feed-forward baselines.

Critical strengths
  1. Important problem. The paper addresses motion blur in feed-forward 3D reconstruction, a practical challenge for reconstruction from real-world captures.

  2. Effective motion-aware design. Explicit trajectory modeling and motion-guided reconstruction improve pose accuracy and rendering quality over the evaluated feed-forward baselines on both synthetic and real blur.

Critical weaknesses
  1. Limited comparison with cascaded baselines. Comparing only against NAFNet + WorldMirror leaves unclear whether the gains stem from the proposed framework or limitations of the deblurring frontend. Additional cascaded baselines using Restormer or FMA-Net, under matched input conditions, would strengthen the comparison.

  2. Unclear training parity with the fine-tuned baseline. Although fine-tuned WorldMirror uses the same blurred training data and training order, it is unclear whether it receives the same sharp-image supervision and training budget as MoB3R. This makes it difficult to isolate the contribution of the proposed architecture from differences in training.

  3. Dependence on sharp-image supervision. The ablation study does not evaluate removing the sharp-image loss. Such an experiment would clarify how much reconstruction quality depends on paired sharp targets and how effective motion guidance and reblur consistency are without this supervision.

  4. Missing NVS ablations. Table 5 reports pose accuracy and input-view deblurring quality, but omits held-out novel-view synthesis results. Since sharp novel-view rendering is a central objective, these results are needed to establish whether each component improves reconstruction beyond the observed viewpoints.

Questions for authors
  1. How are the ground-truth start and end camera poses obtained for trajectory supervision, particularly at timestamps generated by video frame interpolation? Are these poses interpolated from the original frame poses or estimated independently?

  2. Could you provide the exact input and held-out test-view indices for each ExBluRF scene under the 4-, 12-, and 32-view settings, including the hold-out interval? Performance may be sensitive to view selection, particularly with sparse inputs, so these details are important for reproducibility.

Reviewer_VczFOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses feed-forward 3D Gaussian reconstruction from unposed, motion-blurred multi-view images. Building on WorldMirror, the authors first show that an existing single-pose feed-forward model, even after fine-tuning on blurred data, exhibits a systematic pose bias aligned with the underlying blur direction. They propose MoB3R: (1) Additional motion tokens per view to predict the exposure start/end poses, together with the original midpoint pose to give a SLERP exposure trajectory, trained with a reversal-invariant loss to handle forward/backward ambiguity. (2) A per-pixel exposure motion field obtained by projecting the predicted exposure trajectory through predicted depth, which then conditions the Gaussian branch via cross-attention. (3) A reblur cycle loss that averages sharp renderings along the predicted trajectory and compares with the original blurry input. Training uses synthetic blur with a progressive geometry-then-appearance schedule. Evaluation on DL3DV, ExBluRF and DeblurNeRF with 4/12/32 input views reports input-view deblurring, held-out NVS, and relative pose accuracy. MoB3R outperforms WorldMirror, NAFNet+WorldMirror and a fine-tuned WorldMirror baseline on all metrics.

Critical strengths
  1. A clear and testable diagnosis paired with a targeted fix. The motion-aligned pose bias is a good finding, and the trajectory tokens plus reversal-invariant loss are the minimal architectural change that addresses it. Tables 5 and 10 confirm that pose accuracy jumps when the reversal-invariant loss is added, and the bias correlation drops, which closes the loop between analysis and method. The gain also transfers to real blur (ExBluRF) on which the model was never trained on. This is the consideration that supports acceptance: the paper identifies why that fails and then fixes it.

  2. Thorough validation of their claim/model against the right baseline. The key control, fine-tuned WorldMirror with the same data and the same two-stage schedule is present, and MoB3R improves on it consistently across 3 datasets and 3 view counts on deblurring, NVS, and pose. Additionally, progressive-vs-joint, LoRA-vs-full and S-sensitivity ablations are also included.

Critical weaknesses

Modest novelty. Each component is a known idea transplanted into WorldMirror: extra camera tokens, a SLERP trajectory as in BAD-Gaussians, a reblur loss as in BAD-NeRF, cross-attention on a flow map.

Questions for authors
  1. For all models at 4/12/32 views, which poses and point cloud were used for initialization? COLMAP on the blurred subset or dataset-provided sharp-frame poses?
  2. How does MoB3R behave on blur-free inputs compared with the original WorldMirror? Does your trajectory modeling degrade performance when there is no blur?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.