3D Gaussian Splatting & Novel View Synthesis
Dynamic 3D Gaussian Splatting Under Non-Smooth and Discontinuous Motion Dynamics
Generated automatically from the limitations stated in 4 papers (ICLR, CVPR), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
Current dynamic novel view synthesis and 4D Gaussian Splatting frameworks assume smooth, continuous spatio-temporal deformations, linear velocities, and fixed temporal windows. When scenes exhibit abrupt impacts, rapid accelerations, chaotic object trajectories, or zooming and variable frame rates, these continuous deformation priors fail and yield severe geometric tearing, blurring, or total tracking collapse. This fundamentally blocks dynamic radiance fields from capturing real-world interactions such as sports action, mechanical impacts, and sudden dynamic occlusions.
Why it matters
Enables robust free-viewpoint rendering and 4D scene reconstruction for chaotic, fast-action, and discontinuous dynamic phenomena without requiring slow-motion capture or artificial motion smoothing.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Piecewise and keyframe-anchored trajectory parameterization: Replace global low-frequency continuous MLPs or polynomials with adaptive piecewise splines that detect and fit trajectory inflection points, measuring novel-view PSNR and LPIPS on benchmark sequences containing sudden motion transitions.
- 2
Impulse- and acceleration-aware trajectory regularizers: Integrate higher-order temporal derivatives and change-point detection into the Gaussian deformation optimization pipeline, measuring rendering fidelity and point drift during abrupt stops, rebounds, and directional reversals.
- 3
Hybrid multi-rate / asynchronous temporal anchoring: Ingest variable-rate frame inputs or auxiliary motion signals (e.g., optical flow discontinuities or event camera streams) to constrain Gaussian positions across sharp transitions, measuring reconstruction accuracy across variable-fps captures.
Have a different approach?
Describe how you would tackle this problem and we'll look for papers that already do it. Free; your text stays private.
Why it might fail
Severe motion blur in standard camera captures during sudden dynamic transitions may destroy the high-frequency visual features needed to initialize and position Gaussians accurately, regardless of how expressive the temporal trajectory formulation is.
Evidence
Each paper's own statement of the limitation, verbatim.
- UFO-4D: Unposed Feedforward 4D reconstruction from Two ImagesICLR 2026
Assumes constant velocity linear motion and constant photometric brightness, degrading performance over larger time intervals or non-linear dynamic scenes.
- ODE-GS: Latent ODEs for Dynamic Scene Extrapolation with 3D Gaussian SplattingICLR 2026
Assumes smooth and continuous motion dynamics, making it prone to failure on sudden, discontinuous, or chaotic scene transitions.
- AeroGS: Scale-Aware Gaussian Splatting for Pose-Free Dynamic UAV Scene ReconstructionCVPR 2026
Assumes relatively smooth motion; may underperform under extremely fast dynamics or highly cluttered scenes
- 3D Multi-frame Fusion for Video StabilizationCVPR 2024
Relies on fixed intrinsics shared across frames and fixed temporal window of 13 frames, which may not suit varying frame rates or zooming cameras
Nearest existing work
- GaussianVideo: Efficient Video Representation via Hierarchical Gaussian SplattingICCV 2025
- RetimeGS: Continuous-Time Reconstruction of 4D Gaussian SplattingCVPR 2026
- Instant4D: 4D Gaussian Splatting in MinutesNeurIPS 2025
- FastEventDGS: Deformable Gaussian Splatting for Fast Dynamic Scenes from a Single Event CameraCVPR 2026
- 4C4D: 4 Camera 4D Gaussian SplattingCVPR 2026
- Learning Explicit Continuous Motion Representation for Dynamic Gaussian Splatting from Monocular VideosCVPR 2026
- Spacetime Gaussian Feature Splatting for Real-Time Dynamic View SynthesisCVPR 2024
- Learnable Infinite Taylor Gaussian for Dynamic View RenderingCVPR 2025
- A Compact Dynamic 3D Gaussian Representation for Real-Time Dynamic View SynthesisECCV 2024
- EventSplat: 3D Gaussian Splatting from Moving Event Cameras for Real-time RenderingCVPR 2025
- Kinematics-Driven Gaussian Shape Deformation for Blurry Monocular Dynamic ScenesICML 2026
- SC-GS: Sparse-Controlled Gaussian Splatting for Editable Dynamic ScenesCVPR 2024
- Gaussian-Flow: 4D Reconstruction with Dynamic 3D Gaussian ParticleCVPR 2024
- 4D Gaussian Splatting in the Wild with Uncertainty-Aware RegularizationNeurIPS 2024
- 4D3R: Motion-Aware Neural Reconstruction and Rendering of Dynamic Scenes from Monocular VideosNeurIPS 2025
Related open problems
3D Gaussian Splatting & Novel View Synthesis
Cross-Domain Robustness and Generalization Benchmarking for 3D Gaussian Splatting
Novel view synthesis and 3D Gaussian Splatting methods are currently evaluated almost exclusively on room-scale indoor datasets such as Replica, ScanNet++, and Gibson. Because no single pipeline has been systematically evaluated across indoor, outdoor unbounded, panoramic, and object-centric domains, the robustness and failure modes of these representations under domain shift remain entirely unknown. Practitioners and researchers cannot determine whether observed performance gains are artifacts of bounded indoor geometries or if current densification and optimization heuristics translate to unconstrained environments.
3D Gaussian Splatting & Novel View Synthesis
Robust Wide-Baseline 3D Gaussian Splatting Under Unreliable Foundation-Model Geometry Priors
Modern feed-forward 3D Gaussian Splatting pipelines rely on pairwise foundation models (such as MASt3R or VGGT) to supply initial coordinate pointmaps and relative camera poses. Because these foundation predictors degrade significantly as angular view differences increase and inject persistent geometric noise, downstream view synthesis fails when views are not captured in dense, small-step sequences. Treating these imperfect upstream predictions as rigid pseudo-ground truth structurally blocks feed-forward novel view synthesis on sparse, wide-baseline image collections.