acceptodds
Under review as a conference paper at ICLR 2027

NBField: Neural Burst Field for Full-Resolution Blind Deblurring in the Wild

Abstract

Burst deblurring seeks to recover a single sharp photograph from multiple camera-shaken frames. Yet many conventional methods rely on translation-based, spatially invariant blur models, whereas real camera motion—from handheld photography to imaging on moving platforms—is often rotation-dominated and strongly spatially varying at high resolution. In this work, we reinterpret burst reconstruction in the domain of camera rotation, representing exposure motion in SO(3) and sharing a single 2D implicit neural representation across all frames. A differentiable forward model integrates the neural image along rotational homographies, while alternating optimization jointly refines the image and gyroscope-initialized motion without the need for supervised training. Across 447 in-the-wild GoPro bursts spanning severe shake blur and both natural- and low-light scenes, our method reconstructs native 12 megapixels (12MP) images and improves PSNR by an average of 5.75 dB over the best blurry input frame, and 3.87 dB over the best baseline. Each burst is reconstructed in 2.4 minutes on a single GPU, demonstrating practical full-resolution deblurring under everyday capture conditions.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
burst deblurring, blind reconstruction, full resolution, neural field, low light imaging

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_RYJpOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper formulates burst deblurring as a 2D/SO(3) inverse problem: each blurry frame is modeled as the exposure-time integral of a shared hash-grid INR warped along rotational homographies. Motion is initialized from gyro data and refined with a Catmull–Rom spline on SO(3), and image and motion are optimized alternately without pretraining.

Critical strengths
  • A physically grounded, depth-independent rotational model captures spatially varying blur while staying in 2D, avoiding NeRF-style rendering costs.
  • Using gyro data as initialization, with a shared low-dimensional spline correction, is principled and stable.
  • Pixel-query sampling makes test-time optimization practical at 12MP.
  • The large real-world dataset and the stratified motion/luminance analysis are convincing, especially the gains under severe motion.
Critical weaknesses
  • The main text does not describe how ground truth was obtained for handheld captures, which is critical to all reported results.
  • No ablations isolate the contributions of gyro initialization, the spline parameterization, or the loss weighting. Since the individual components are established, the value of their integration needs empirical support.
  • Some baselines (PI-CNN, Deblur-INR) operate far outside their intended regimes, and input conditions across methods are unclear.
Questions for authors
  • How was ground truth captured and registered?
  • How does performance change with fixed gyro motion or without gyro initialization?
  • How robust is the method to gyro–camera synchronization and calibration errors?
  • Do the authors plan to publicly release the dataset, including gyro data and ground truth?
4 comments
AuthorsOpenReview ↗

[1/2] Response to Reviewer RYJp

W3. Some baselines (PI-CNN, Deblur-INR) operate far outside their intended regimes, and input conditions across methods are unclear.

With our understanding, PI-CNN is directly relevant to our task. PI-CNN was proposed to reconstruct one sharp image from an arbitrary-length burst under realistic camera shake. We use the authors' released ImageNet checkpoint from https://people.csail.mit.edu/miika/eccv18_deblur/ with our ten blurry frames and sliding-window reconstruction. No gyro information is supplied.

We include Deblur-INR as a related deblurring method that uses neural field. We acknowledge that its single-image, spatially-invariant convolution representation differs from our SO(3) representation. We first select the blurry frame with the highest post-registration PSNR, then feed that frame to Deblur-INR for sliding-window reconstruction. We have articulated these differences in the manuscript. More baseline implementation details can be found in Supplementary Materials, Section C.

Q1. How was ground truth captured and registered?

We thank the reviewer for pointing out this omission. For each capture, we take a 3-4 seconds video at 8K, 30 FPS with a handheld GoPro Mission 1. Each frame has a 33 ms exposure. Before the video starts, there is a 3s count down, during which we hold our hands still. We continue to stabilize our hands into the first ~1s of the capture, whereas in the remaining time we randomly tremor our hands to create realistic motion. Since there’s no interruption between the 3s delay capture and the first ~1s second, we avoided small motion created by clicking the camera shutter.

The first frame of the ~1s still video is used as the sharp references/ground truth. From the remaining blurry, hand-tremor videos (excluding the first ~1s), we select a 330ms burst with the greatest cumulative gyroscope motion as our measurements. Each burst consists of ten consecutive blurry frames where each blurry frame is a single 33 ms exposure. Upon visual inspection, 447 ground truth images out of 448 captures are clearly without motion artifacts. We removed the one case where the ground truth image is blurry. We do not add extra noise to the blurry images, since all images are real GoPro captures and already include real photon shot noise and read noise. We also do not perform any additional black-level subtraction.

We have added a description of how the dataset is generated in Section 4.1.

Q3. How robust is the method to gyro–camera synchronization and calibration errors?

We emphasize that NBField uses built-in gyroscope measurements from a commercial GoPro camera, without requiring specialized sensing or synchronization hardware. We extract timestamped gyroscope data using the official GoPro GPMF format, with a median sampling rate of 197.5 Hz. We calibrate a fixed gyro-to-camera rotation and a 7.925 ms temporal offset to establish the correct timing reference. Both calibration parameters remain fixed throughout reconstruction.

Importantly, even with accurate calibration and synchronization, gyroscope-derived motion remains imperfect for native 12MP deblurring, where small motion errors can cause visible artifacts. This motivates our image-driven spline refinement, which improves PSNR by 0.99–2.57 dB over the fixed-gyro ablation. See Table 2 in the revised manuscript.

These results demonstrate that NBField achieves high-quality 12MP reconstruction using readily available commercial inertial measurements, while compensating for residual motion-estimation errors through image-driven refinement.

Q4. Do the authors plan to publicly release the dataset, including gyro data and ground truth?

Yes. We plan to publicly release the dataset, including the measured blurry bursts, synchronized gyroscope data, and sharp ground-truth references. The Supplementary Materials also include additional representative reconstruction examples.

AuthorsOpenReview ↗

[2/2] Response to Reviewer RYJp

Q2. How does performance change with fixed gyro motion or without gyro initialization?

We would like to clarify that Supplementary Material Section D, Table 5 reports ablation studies regarding gyroscope initialization and motion refinement. We additionally conducted two new ablation studies to isolate the contribution of the neural image representation and the brightness weighting.

Specifically, we consider four ablation variants:

  • No gyro: We remove gyroscope initialization and instead randomly initialize the motion trajectory, which is subsequently optimized using the spline.
  • No spline: We use the gyroscope-derived motion trajectory without spline refinement (ϕ=0)(\phi = 0), while keeping the image representation, forward model, and optimization schedule unchanged.
  • No neural field: We replace the neural field with an RGB pixel grid using differentiable bilinear interpolation, while retaining the same motion model and reconstruction framework.
  • No brightness weighting: We remove the brightness weighting in equation (5) by setting wiw_i to 1.

The following table summarizes the numerical results:

MethodOverall PSNR ↑SSIM ↑LPIPS ↓Time (min) ↓Mem. (GiB) ↓Severe PSNR ↑SSIM ↑LPIPS ↓Low-light PSNR ↑SSIM ↑LPIPS ↓
NBField (ours)39.410.950.292.45.532.260.920.3042.450.970.31
No neural field32.090.780.512.55.827.970.780.5433.170.760.50
No gyro28.830.890.393.75.419.750.770.5032.930.930.37
No spline37.850.940.312.55.229.670.890.3441.260.960.33
No brightness weighting37.410.940.332.35.530.890.910.3340.050.950.35

As shown in the table above, NBField achieves over 10 dB improvement in PSNR compared to the no-gyro variant, demonstrating the importance of accurate motion initialization. Compared to the no-spline variant, NBField achieves a 1.56 dB improvements overall and 2.59 dB improvement for severe motion, demonstrating that motion refinement with spline provides additional reconstruction benefits.

Furthermore, NBField outperforms the no-neural-field variant by more than 7 dB overall and 9.28 dB under low-light conditions. This demonstrates that the neural image representation plays a substantial role in reconstruction quality, beyond the benefits of the physical forward model and motion refinement. NBField also achieves over 2 dB improvement compared to the no-brightness-weighting ablation. All ablation variants have similar time and memory costs as NBField. We have incorporated these ablation results and a corresponding discussion into Section 4.3 of the revised manuscript.

Dear Reviewer RYJp, thank you again for your time and feedback. Could you let us know whether our response sufficiently addressed your concerns? If there are any remaining points that we could clarify, we are happy to discuss further.

Thank you for the authors' response. I believe that their response address most of my concerns.

Therefore, I raise my score from WR to WA.

Reviewer_ZxdhOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper reframes burst deblurring as a 2D/SO(3) inverse problem, which jointly optimizes a shared 2D implicit neural representation and a gyroscope-initialized camera rotation trajectory.

Critical strengths

This paper combines 2D implicit neural representations with SO(3)SO(3) rotational motion modeling to achieve unsupervised blind deblurring for high-resolution real-world scenes.

Critical weaknesses
  1. The paper lacks comprehensive ablation studies and analytical experiments. Although the method performs well, the authors do not explain the underlying reasons for its success.
  2. I wonder if the proposed method only applies to static scenes under SO(3) camera motion. How does it perform in scenarios involving moving objects or forward/backward camera translations?
Questions for authors

The lack of analysis on the proposed method led me to lower my score.

3 comments
AuthorsOpenReview ↗

Response to Reviewer Zxdh

  1. The paper lacks comprehensive ablation studies and analytical experiments. Although the method performs well, the authors do not explain the underlying reasons for its success.

    We would like to clarify that Supplementary Material Section D, Table 5 reports ablation studies regarding gyroscope initialization and motion refinement. We additionally conducted two new ablation studies to isolate the contribution of the neural image representation and the brightness weighting.

    Specifically, we consider four ablation variants:

    • No gyro: We remove gyroscope initialization and instead randomly initialize the motion trajectory, which is subsequently optimized using the spline.
    • No spline: We use the gyroscope-derived motion trajectory without spline refinement (ϕ=0)(\phi = 0), while keeping the image representation, forward model, and optimization schedule unchanged.
    • No neural field: We replace the neural field with an RGB pixel grid using differentiable bilinear interpolation, while retaining the same motion model and reconstruction framework.
    • No brightness weighting: We remove the brightness weighting in equation (5) by setting wiw_i to 1.

    The following table summarizes the numerical results:

    MethodOverall PSNR ↑SSIM ↑LPIPS ↓Time (min) ↓Mem. (GiB) ↓Severe PSNR ↑SSIM ↑LPIPS ↓Low-light PSNR ↑SSIM ↑LPIPS ↓
    NBField (ours)39.410.950.292.45.532.260.920.3042.450.970.31
    No neural field32.090.780.512.55.827.970.780.5433.170.760.50
    No gyro28.830.890.393.75.419.750.770.5032.930.930.37
    No spline37.850.940.312.55.229.670.890.3441.260.960.33
    No brightness weighting37.410.940.332.35.530.890.910.3340.050.950.35

    As shown in the table above, NBField achieves over 10 dB improvement in PSNR compared to the no-gyro variant, demonstrating the importance of accurate motion initialization. Compared to the no-spline variant, NBField achieves a 1.56 dB improvements overall and 2.59 dB improvement for severe motion, demonstrating that motion refinement with spline provides additional reconstruction benefits.

    Furthermore, NBField outperforms the no-neural-field variant by more than 7 dB overall and 9.28 dB under low-light conditions. This demonstrates that the neural image representation plays a substantial role in reconstruction quality, beyond the benefits of the physical forward model and motion refinement. NBField also achieves over 2 dB improvement compared to the no-brightness-weighting ablation. All ablation variants have similar time and memory costs as NBField. We have incorporated these ablation results and a corresponding discussion into Section 4.3 of the revised manuscript.

  2. I wonder if the proposed method only applies to static scenes under SO(3) camera motion. How does it perform in scenarios involving moving objects or forward/backward camera translations?

    We acknowledge that the current NBField formulation primarily targets static scenes under rotation-dominated camera motion. This assumption is well motivated by real-world handheld photography, where rotational camera shake is often the dominant source of motion blur, particularly for outdoor scenes with relatively distant objects [1].

    Our SO(3) formulation therefore captures an important class of real-world camera shake without introducing additional unknowns associated with scene depth and 3D geometry. We acknowledge that this approximation becomes less accurate for close-range scenes or captures involving substantial camera translation.

    Nevertheless, we provide an illustrative example in Supplementary Material Section E, Figure 7, featuring Barkley, a moving dog captured at approximately one meter from the camera. Despite the dog's slight movements during the burst, NBField produces a visually sharp reconstruction with clearly resolved fur textures and object boundaries. This example suggests that NBField can tolerate moderate deviations from the static-scene assumption while maintaining high reconstruction quality.

    [1] Oliver Whyte, Josef Sivic, Andrew Zisserman, and Jean Ponce. Non-uniform deblurring for shaken images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 491–498, 2010.

Dear Reviewer Zxdh, thank you again for your time and feedback. Could you let us know whether our response sufficiently addressed your concerns? If there are any remaining points that we could clarify, we are happy to discuss further.

I raise my score from 2 to 3.

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

This paper studies full-resolution, gyro-assisted burst deblurring for static scenes where camera motion is mainly rotational. NBField represents the sharp image with a shared 2D neural image field using multiresolution hash encoding. Each blurry frame is reconstructed by integrating this field along the camera motion during exposure, in linear RGB. The gyro trajectory stays fixed during optimization. A learned Catmull-Rom residual is mapped through the SO(3) exponential and composed with this trajectory. The image and residual motion are updated in alternating steps, with stochastic pixel sampling, a brightness-weighted Huber loss, and a progressive optimization schedule. The main contribution is to put the exposure model, neural image fitting, and gyro-based motion correction together in a form that can run at 12MP. On the reported dataset of 447 ten-frame handheld bursts, the method obtains 39.41 dB average PSNR. This is 3.87 dB above GyroDVD overall and 6.47 dB above it on the severe-motion subset.

Critical strengths
  • Using a shared 2D image field with continuous rotational exposure integration is a reasonable choice here. Under the assumed motion, it can model spatially varying blur without estimating scene depth or using volumetric rendering. One continuous trajectory explains both the motion between frames and the blur within each exposure. For me, this is an important strength because the model directly addresses the difficult part of recovering a high-resolution image from a small burst with strong motion blur. The authors keep the problem relatively compact for the capture conditions they target, and this makes the optimization manageable.

  • Table 1 reports 39.41 dB for NBField versus 35.54 dB for GyroDVD overall. On the severe motion subset, the numbers are 32.26 and 25.79 dB. GyroDVD is the strongest baseline in the table, and it also uses gyro information. So the comparison is more informative than comparing only with methods that do not have sensor input. The larger gain under severe motion is also directly relevant to the motivation of this paper.

  • The reported 2.4 minute runtime and 5.5 GiB memory use at 12MP make the method usable for offline reconstruction on the evaluated hardware. These results are good evidence for the complete system. They do not yet tell us how much each component contributes.

Critical weaknesses
  • An important missing comparison in the main text is the full method versus a version that fixes the trajectory to the integrated gyro measurements, with phi = 0. This version should still optimize the same image field using the same observations and exposure model. Without this comparison, it is difficult to know how much of the gain comes from residual motion estimation, and how much can already be obtained by fitting the image with fixed-gyro exposure integration.

  • The main text describes the handheld captures and reports full-reference metrics, but I could not find enough explanation of how the sharp reference images were obtained or constructed. We need to know the relation between the blurry observations and the references, their noise and radiometric processing, and the registration used for evaluation.

Questions for authors
  • What do motion refinement and the neural image representation add? Could you provide a controlled comparison between full NBField and a fixed-gyro version with phi = 0, or point to where this comparison is already reported? Please keep the observations, image representation, and exposure model matched. Results with the compute cost would be useful.

  • How are the sharp references obtained, and how are the metrics computed? Please explain how the sharp reference for each burst is acquired or constructed, including its temporal and photometric relation to the blurry observations. Are the blurry frames actual single-exposure captures, or are they constructed from shorter exposures? What processing is applied to the inputs and references, including linearization, black-level treatment, noise handling, and radiometric mapping?

  • Please also specify the evaluation color space, the registration transform allowed, and the valid-pixel mask or crop. Are these applied in the same way across methods? The discussion of dark-region effects on PSNR is helpful, but the evaluation protocol still needs to be described.

4 comments
AuthorsOpenReview ↗

[1/2] Response to Reviewer uxeA

  1. What do motion refinement and the neural image representation add? Could you provide a controlled comparison between full NBField and a fixed-gyro version with ϕ=0\phi = 0, or point to where this comparison is already reported? Please keep the observations, image representation, and exposure model matched. Results with the compute cost would be useful.

    We would like to clarify that Supplementary Material Section D, Table 5 reports ablation studies regarding gyroscope initialization and motion refinement. We additionally conducted two new ablation studies to isolate the contribution of the neural image representation and the brightness weighting.

    Specifically, we consider four ablation variants:

    • No gyro: We remove gyroscope initialization and instead randomly initialize the motion trajectory, which is subsequently optimized using the spline.
    • No spline: We use the gyroscope-derived motion trajectory without spline refinement (ϕ=0)(\phi = 0), while keeping the image representation, forward model, and optimization schedule unchanged.
    • No neural field: We replace the neural field with an RGB pixel grid using differentiable bilinear interpolation, while retaining the same motion model and reconstruction framework.
    • No brightness weighting: We remove the brightness weighting in equation (5) by setting wiw_i to 1.

    The following table summarizes the numerical results:

    MethodOverall PSNR ↑SSIM ↑LPIPS ↓Time (min) ↓Mem. (GiB) ↓Severe PSNR ↑SSIM ↑LPIPS ↓Low-light PSNR ↑SSIM ↑LPIPS ↓
    NBField (ours)39.410.950.292.45.532.260.920.3042.450.970.31
    No neural field32.090.780.512.55.827.970.780.5433.170.760.50
    No gyro28.830.890.393.75.419.750.770.5032.930.930.37
    No spline37.850.940.312.55.229.670.890.3441.260.960.33
    No brightness weighting37.410.940.332.35.530.890.910.3340.050.950.35

    As shown in the table above, NBField achieves over 10 dB improvement in PSNR compared to the no-gyro variant, demonstrating the importance of accurate motion initialization. Compared to the no-spline variant, NBField achieves a 1.56 dB improvements overall and 2.59 dB improvement for severe motion, demonstrating that motion refinement with spline provides additional reconstruction benefits.

    Furthermore, NBField outperforms the no-neural-field variant by more than 7 dB overall and 9.28 dB under low-light conditions. This demonstrates that the neural image representation plays a substantial role in reconstruction quality, beyond the benefits of the physical forward model and motion refinement. NBField also achieves over 2 dB improvement compared to the no-brightness-weighting ablation. All ablation variants have similar time and memory costs as NBField. We have incorporated these ablation results and a corresponding discussion into Section 4.3 of the revised manuscript.

AuthorsOpenReview ↗

[2/2] Response to Reviewer uxeA

  1. How are the sharp references obtained, and how are the metrics computed? Please explain how the sharp reference for each burst is acquired or constructed, including its temporal and photometric relation to the blurry observations. Are the blurry frames actual single-exposure captures, or are they constructed from shorter exposures? What processing is applied to the inputs and references, including linearization, black-level treatment, noise handling, and radiometric mapping?

    We thank the reviewer for pointing out this omission. For each capture, we take a 3-4 seconds video at 8K, 30 FPS with a handheld GoPro Mission 1. Each frame has a 33 ms exposure. Before the video starts, there is a 3s count down, during which we hold our hands still. We continue to stabilize our hands into the first ~1s of the capture, whereas in the remaining time we randomly tremor our hands to create realistic motion. Since there’s no interruption between the 3s delay capture and the first ~1s second, we avoided small motion created by clicking the camera shutter.

    The first frame of the ~1s still video is used as the sharp references/ground truth. From the remaining blurry, hand-tremor videos (excluding the first ~1s), we select a 330ms burst with the greatest cumulative gyroscope motion as our measurements. Each burst consists of ten consecutive blurry frames where each blurry frame is a single 33 ms exposure. Upon visual inspection, 447 ground truth images out of 448 captures are clearly without motion artifacts. We removed the one case where the ground truth image is blurry. We do not add extra noise to the blurry images, since all images are real GoPro captures and already include real photon shot noise and read noise. We also do not perform any additional black-level subtraction.

    Our radiometry and evaluation scheme can be found in the Supplementary Materials. As described in Section B, we use 10-bit GP-Log2 for all video captures to maximize the input dynamic range. We then convert the log values to linear Rec.2020 for reconstruction, and subsequently to Rec.709 (natural color) for metric computation.

    We have added a description of how the dataset is generated in Section 4.1.

  2. Please also specify the evaluation color space, the registration transform allowed, and the valid-pixel mask or crop. Are these applied in the same way across methods? The discussion of dark-region effects on PSNR is helpful, but the evaluation protocol still needs to be described.

    These information can be found in Section B in Supplemental Materials. The evaluation color space is Rec.709, as described above. We estimate a global projective homography between the reconstruction and ground truth using SIFT and MAGSAC before metric computation.

    Our reported metrics are computed over the central 4000×30004000 \times 3000 ROI in the 5000×40005000 \times 4000 reconstruction canvas (with dimensions swapped for portrait captures).

    As described in Section C in Supplemental Materials, the same evaluation procedure is applied to NBField and all baselines. If their reconstructions fall outside of the ground truth ROI after registration, we exclude the invalid pixels from metrics computations. There is no additional masking or cropping during evaluation.

Dear Reviewer uxeA, thank you again for your time and feedback. Could you let us know whether our response sufficiently addressed your concerns? If there are any remaining points that we could clarify, we are happy to discuss further.

Thank you for the detailed response and additional experiments. The comparison with the various ablation variants clarifies the contributions of motion refinement and the neural image representation. The explanation of reference capture, image processing, and the common evaluation procedure also addresses my concerns about the dataset and reported metrics. My main concerns are now resolved, and I will keep my rating of weak acceptance.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We thank the reviewers for their constructive feedback. In response, we have revised the manuscript to strengthen the experimental analysis and clarify the data acquisition and evaluation protocols. The major changes are summarized below:

  • Expanded ablation studies (Section 4.3, Table 2). We have incorporated four ablation studies into Table 2, examining the contributions of gyroscope initialization, spline-based motion refinement, neural image representation, and brightness-weighted loss. We added a corresponding discussion as Section 4.3 to explain how each component contributes to reconstruction quality, particularly under severe motion and low-light conditions.

  • Clarified ground-truth acquisition (Section 4.1). We have added a detailed description of our real-world data collection procedure, explaining how sharp reference images and blurry bursts are acquired. We also clarify that each blurry frame is an actual single exposure rather than synthetically generated blur.

  • Clarified evaluation and baseline protocols (Supplementary Materials, Sections B–C). We have clarified the radiometric processing, registration procedure, and evaluation region. We also provide additional explanations of baseline input conditions and implementation choices to facilitate fair comparison and reproducibility.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.