Research map / 3D vision / NeRF and radiance fields

DynaMoDe-NeRF: Motion-aware Deblurring Neural Radiance Field for Dynamic Scenes

CVPR 2025 · 3D vision › NeRF and radiance fields · approach: Few-shot neural rendering architectures · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DynaMoDe-NeRF, a self-supervised NeRF that renders motion-blur-free novel views of dynamic scenes from multi-view blurry videos with object motion blur, while recovering 3D object velocity and trajectory.

Problem

Novel view synthesis of dynamic scenes from multi-view videos corrupted by local object motion blur, where object velocity is unknown and time-varying; prior NeRF deblurring work only handles static scenes with camera or defocus blur.

Approach

Relates 3D object motion to continuous blur kernel locations across views and time via an analytical projection model, learns kernel weights with a small depth- and velocity-aware MLP (BlurKernelNet), and jointly optimizes a HexPlane scene, control velocities, and blur parameters with motion regularizers.

Key results

• On the Cube scene, the method reaches 32.12 PSNR / 0.94 SSIM in static regions and 26.93 PSNR in dynamic regions, the best among compared methods in the dynamic region of that scene.
• Across the four synthetic scenes, the overall PSNR of DynaMoDe-NeRF is 31.47, 32.59, 37.64 and 34.78, but 4DGS scores higher on Lychee-type overall PSNR (35.29) and Sphere-type scores vary, so gains are not uniform.
• Recovered speed and 3D trajectory have cosine similarity above 0.81 to ground truth, with speed above 0.98 on the Cube example.

Measured effects reported in the paper

Surprising findings

• 4DGS obtains higher overall PSNR than the proposed method on the Lychee scene (35.29 vs 32.59), so the method does not dominate every metric.

Limitations

• Requires an initial rough motion mask of the moving foreground.
• Assumes the moving object is rigid and translational motion only.
• Handling very severe blur remains a challenge.

Experimental setup

Datasets: Deblur-NeRF scenes (adapted) Blender synthetic blurred multi-view videos Real GoPro multi-view captures

Benchmarks: DynaMoDe-NeRF synthetic dataset (Cube, Lychee, Monkey, Sphere)

Metrics: PSNR SSIM LPIPS speed and trajectory cosine similarity

Baselines: HexPlane MixVoxel K-Planes 4DGS (GS4D) Deblur-NeRF DP-NeRF DyBluRF

Builds on: HexPlane NeRF Deblur-NeRF DP-NeRF

Closest papers by meaning

• Mitigating Motion Blur in Neural Radiance Fields with Events and Frames (CVPR 2024 · similarity 0.91)
• DyBluRF: Dynamic Neural Radiance Fields from Blurry Monocular Video (CVPR 2024 · similarity 0.90)
• LuSh-NeRF: Lighting up and Sharpening NeRFs for Low-light Scenes (NeurIPS 2024 · similarity 0.89)
• NeRF On-the-go: Exploiting Uncertainty for Distractor-free NeRFs in the Wild (CVPR 2024 · similarity 0.89)
• Joint Optimization of Neural Radiance Fields and Continuous Camera Motion from a Monocular Video (CVPR 2025 · similarity 0.88)
• Gear-NeRF: Free-Viewpoint Rendering and Tracking with Motion-aware Spatio-Temporal Sampling (CVPR 2024 · similarity 0.88)