Research map / Point clouds / Point cloud classification
TSDF-Based Efficient Motion-Compensated Temporal Interpolation for 3D Dynamic Sequences
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 a TSDF-volume-based, RAFT-derived bi-directional motion-compensated network for real-time temporal interpolation of 3D dynamic sequences, outperforming point-cloud, NeRF and Gaussian Splatting baselines.
Problem
Temporal frame interpolation for 3D dynamic sequences (meshes/3D scans) to raise frame rates for VR and displays, where prior point-cloud, NeRF and Gaussian-splatting methods suffer from irregular point density, spurious errors, slow per-scene optimization, or overfitting.
Approach
Represent each frame as a dense TSDF voxel grid, build bi-directional 6D correlation volumes and GRU-based RAFT-style flow refinement in 3D, estimate intermediate flows with a learned network conditioned on time, and merge forward- and backward-warped TSDF volumes adaptively.
Key results
- For x3 frame-rate conversion, overall CD 0.043 vs 0.080 for NeuralPCI (CD in ×10^-3, lower is better); EMD 6.849 vs 12.338 overall.
- PSNR overall 36.76 on 4DHO and 32.76 on MPEG, versus 30.53 and 30.59 for Deformable-GS (and 16.25 and 19.85 for D-NeRF).
- Inference takes 0.285-0.338 s per frame vs 0.618 s for NeuralPCI and 8.34 s for D-NeRF; Deformable-GS is faster at 0.122 s per frame.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TSDF-based interpolation instead of NeuralPCI | 3D dynamic sequence interpolation · MPEG | CD: 0.095 -> 0.089 | Helped |
| Use TSDF-based interpolation instead of NeuralPCI | 3D dynamic sequence interpolation · MPEG | EMD: 9.502 -> 10.456 | Hurt |
| Use TSDF-based interpolation instead of NeuralPCI | 3D dynamic sequence interpolation · MITAMA | CD: 0.066 -> 0.026 | Helped |
| Use TSDF-based interpolation instead of NeuralPCI | 3D dynamic sequence interpolation · MITAMA | EMD: 12.903 -> 5.649 | Helped |
| Use TSDF-based interpolation instead of NeuralPCI | 3D dynamic sequence interpolation · 4DHO | CD: 0.080 -> 0.016 | Helped |
| Use TSDF-based interpolation instead of NeuralPCI | 3D dynamic sequence interpolation · 4DHO | EMD: 14.608 -> 4.441 | Helped |
| Use TSDF-based interpolation instead of NeuralPCI | 3D dynamic sequence interpolation · Overall | CD: 0.080 -> 0.043 | Helped |
| Use TSDF-based interpolation instead of NeuralPCI | 3D dynamic sequence interpolation · Overall | EMD: 12.338 -> 6.849 | Helped |
| Use TSDF-based interpolation instead of Deformable-GS | 3D dynamic sequence interpolation · 4DHO | PSNR: 30.529 -> 36.762 | Helped |
| Use TSDF-based interpolation instead of Deformable-GS | 3D dynamic sequence interpolation · MPEG | PSNR: 30.588 -> 32.763 | Helped |
| Add learned intermediate motion estimation (ablation) | TSDF-based interpolation · Overall | CD: 0.065 -> 0.043 | Helped |
| Add learned intermediate motion estimation (ablation) | TSDF-based interpolation · Overall | EMD: 8.527 -> 6.849 | Helped |
Surprising findings
- On the MPEG dataset the proposed method is not better than NeuralPCI in EMD (10.456 vs 9.502), so gains are dataset-dependent.
- Ablation shows removing learned motion estimation hurts most (CD rises to 0.065 from 0.043), indicating motion is nonlinear.
Limitations
- Uses only geometric TSDF information for motion, leading to occasional texture temporal inconsistencies.
- Per-frame inference is slower than Deformable-GS.
- Comparisons with point cloud, NeRF and Gaussian methods are across different output forms, making time comparisons difficult.
Experimental setup
Datasets: MPEG dynamic mesh dataset MITAMA 4DHumanOutfit
Benchmarks: MPEG dynamic mesh (Longdress, Loot, Soldier) MITAMA (Swing, Crane, Handstand) 4DHumanOutfit (deb-jea-walk, leo-tig-jump, ray-own-torso, ted-opt-side)
Metrics: Chamfer Distance Earth Mover's Distance PSNR inference time
Baselines: NeuralPCI D-NeRF Deformable-GS
Builds on: RAFT PointINet NeuralPCI
Closest papers by meaning
- LIM: Large Interpolator Model for Dynamic Reconstruction (CVPR 2025 · similarity 0.84)
- HAIF-GS: Hierarchical and Induced Flow-Guided Gaussian Splatting for Dynamic Scene (NeurIPS 2025 · similarity 0.83)
- Degrees of Freedom Matter: Inferring Dynamics from Point Trajectories (CVPR 2024 · similarity 0.82)
- TFS-NeRF: Template-Free NeRF for Semantic 3D Reconstruction of Dynamic Scene (NeurIPS 2024 · similarity 0.82)
- Enhancing 3D Reconstruction for Dynamic Scenes (NeurIPS 2025 · similarity 0.82)
- PAPR in Motion: Seamless Point-level 3D Scene Interpolation (CVPR 2024 · similarity 0.81)
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.