Research map / 3D vision / Gaussian splatting
Feed-Forward Bullet-Time Reconstruction of Dynamic Scenes from Monocular Videos
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.
Introduces BTimer, a feed-forward transformer that reconstructs a frozen 3D Gaussian Splatting scene at any bullet timestamp from monocular context frames, plus a Novel Time Enhancer for interpolating intermediate frames.
Problem
Feed-forward 3D reconstruction and novel view synthesis has been limited to static scenes; dynamic scene reconstruction from monocular video typically needs slow per-scene optimization or is restricted to synthetic object-centric data.
Approach
Add a bullet-time embedding to context frame tokens so a transformer aggregates all context frames into a full 3DGS scene frozen at a target timestamp; this unifies static and dynamic training, enabling pretraining on static data, and a 3D-free NTE module synthesizes intermediate frames to handle fast motion.
Key results
- On NVIDIA Dynamic Scene, BTimer reaches 25.82 PSNR, third among baselines and above Casual-FVS (24.57) with 0.78 s reconstruction time versus hours for optimization methods.
- On DyCheck iPhone, BTimer achieves 16.52 PSNR, 0.570 SSIM and 0.338 LPIPS, ranking second in SSIM and LPIPS and beating PGDVS on all three metrics.
- Feed-forward inference takes 150 ms for 12 context views at 256x256 on a single A100.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use BTimer instead of PGDVS | Dynamic novel view synthesis · DyCheck iPhone | PSNR | Helped |
| Use BTimer instead of PGDVS | Dynamic novel view synthesis · DyCheck iPhone | SSIM | Helped |
| Use BTimer instead of PGDVS | Dynamic novel view synthesis · DyCheck iPhone | LPIPS | Helped |
| Use BTimer instead of per-scene optimization methods | Dynamic novel view synthesis · NVIDIA Dynamic Scene | PSNR | Mixed |
| Add static-scene Stage 1 pretraining (ablation) | BTimer · Dynamic scenes | reconstruction quality | Helped |
| Pretrain on multiple diverse datasets instead of RE10K alone (ablation) | BTimer · Dynamic scenes | reconstruction quality | Helped |
| Co-train on static scenes in Stage 2 (ablation) | BTimer · Dynamic scenes | reconstruction quality | Helped |
| Add interpolation supervision (ablation) | BTimer · Dynamic scene reconstruction | reconstruction quality | Helped |
| Add Novel Time Enhancer (NTE) (ablation) | BTimer · Fast or complex motion | ghosting artifacts | Helped |
| Use NTE for novel views far from the input camera trajectory (ablation) | NTE · Novel views | novel-view synthesis quality | Hurt |
| Increase the number of context frames (ablation) | BTimer · Dynamic scene reconstruction | scene completeness | Helped |
Surprising findings
- Without interpolation supervision the model produces white-edge artifacts because Gaussians are placed close to the camera to cheat the loss.
Limitations
- Recovered geometry and depth are less accurate since the method targets novel view synthesis.
- Pixel-aligned Gaussian representation cannot represent temporal deformations; motion must be recovered via post-processing.
- NTE struggles to synthesize novel views far from the input camera trajectory.
Experimental setup
Datasets: Objaverse RE10K MVImgNet DL3DV Kubric PointOdyssey DynamicReplica Spring PANDA-70M
Benchmarks: DyCheck iPhone NVIDIA Dynamic Scene RealEstate10K Tanks and Temples DAVIS
Metrics: PSNR SSIM LPIPS reconstruction time render FPS
Baselines: TiNeuVox NSFF T-NeRF Nerfies HyperNeRF PGDVS DynNeRF RoDynRF 4D-GS Casual-FVS MVSplat GS-LRM
Builds on: 3D Gaussian Splatting GS-LRM LVSM Plücker embeddings LPIPS
Closest papers by meaning
- 4D3R: Motion-Aware Neural Reconstruction and Rendering of Dynamic Scenes from Monocular Videos (NeurIPS 2025 · similarity 0.88)
- DGS-LRM: Real-Time Deformable 3D Gaussian Reconstruction From Monocular Videos (NeurIPS 2025 · similarity 0.88)
- Instant4D: 4D Gaussian Splatting in Minutes (NeurIPS 2025 · similarity 0.88)
- World from Motion: Generative Dynamic Gaussian Reconstruction from Monocular Video (NeurIPS 2026 · similarity 0.88)
- MoDGS: Dynamic Gaussian Splatting from Casually-captured Monocular Videos with Depth Priors (ICLR 2025 · similarity 0.87)
- Reconstruct, Inpaint, Test-Time Finetune: Dynamic Novel-view Synthesis from Monocular Videos (NeurIPS 2025 · similarity 0.87)
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.