Research map / 3D vision / NeRF and radiance fields
MovingParts: Motion-based 3D Part Discovery in Dynamic Radiance Field
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 MovingParts, a NeRF-based method that jointly models scene motion with Eulerian and Lagrangian modules tied by a cycle-consistency loss, and groups particle trajectories into rigid parts for monocular part discovery and reconstruction.
Problem
Dynamic NeRF methods reconstruct scenes but cannot discover object parts; Eulerian deformation fields make per-object motion hard to extract, and prior part-discovery methods need multi-view input or ellipsoid priors.
Approach
Adds a Lagrangian (particle-trajectory) module alongside the Eulerian deformation field, enforces cycle consistency between them, and applies slot-based hard motion grouping (Gumbel-softmax with straight-through estimator) plus an APE-based group merging step to discover rigid parts from monocular video.
Key results
- On the D-NeRF synthetic dataset, the method averages 32.24 PSNR and 0.97 SSIM, second to TiNeuVox's 32.67 PSNR, while training in about 26 minutes on one RTX3090.
- The method achieves mIOU above 85% on most scenes of its Kubric-generated motion-grouping dataset.
- Wim (Watch-It-Move) averages only 15.72 PSNR on the same benchmark, showing monocular failure of multi-view part-discovery methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| remove temporal enhancement module from TiNeuVox canonical space (ablation) | TiNeuVox · D-NeRF 360 synthetic | PSNR (average): 32.67 -> 31.47 | Hurt |
| Motion-based Lagrangian part grouping with Eulerian-Lagrangian cycle consistency (full MovingParts) | dynamic NeRF · D-NeRF 360 synthetic | SSIM (average): best SSIM 0.97 vs D-NeRF 0.95 | Helped |
| Proposed method vs TiNeuVox for PSNR | MovingParts · D-NeRF 360 synthetic | PSNR (average): 32.24 vs 32.67 | Hurt |
| WIM (Lagrangian multi-view part method) with single-view monocular input | WIM · D-NeRF 360 synthetic | PSNR (average): 15.72 avg PSNR | Hurt |
Surprising findings
- Removing TiNeuVox's temporal enhancement module drops its average PSNR to 31.47, below the proposed method's 32.24.
Limitations
- Motion features are stored in low-dimensional vectors, making very long sequences hard to model; long videos must be split manually.
- Cycle loss applies only to object points, since free-space deformation violates the rigid-motion assumption.
Experimental setup
Datasets: D-NeRF 360-degree synthetic dataset Kubric-generated dataset with GSO objects ManiSkill robotic manipulation sequences HyperNeRF real-world sequences
Benchmarks: D-NeRF 360-degree synthetic dataset Kubric motion grouping dataset
Metrics: PSNR SSIM training time mIOU
Baselines: T-NeRF D-NeRF TiNeuVox NDVG Watch-It-Move
Builds on: NeRF D-NeRF TiNeuVox Watch-It-Move Slot Attention
Closest papers by meaning
- Real-time Photorealistic Dynamic Scene Representation and Rendering with 4D Gaussian Splatting (ICLR 2024 · similarity 0.86)
- Motion Decoupled 3D Gaussian Splatting for Dynamic Object Representation (AAAI 2025 · similarity 0.86)
- Gear-NeRF: Free-Viewpoint Rendering and Tracking with Motion-aware Spatio-Temporal Sampling (CVPR 2024 · similarity 0.86)
- MotionGS: Exploring Explicit Motion Guidance for Deformable 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.86)
- Structure from Collision (CVPR 2025 · similarity 0.85)
- Flux4D: Flow-based Unsupervised 4D Reconstruction (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.