Research map / 3D vision / Gaussian splatting
No Pose, No Problem: Surprisingly Simple 3D Gaussian Splats from Sparse Unposed Images
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.
Feed-forward NoPoSplat predicts 3D Gaussians for all unposed sparse views in the first view's canonical space, trained only with photometric loss, with intrinsic-token scale resolution and a PnP-plus-photometric pose estimator.
Problem
Generalizable feed-forward 3D Gaussian reconstruction from sparse unposed images normally needs known camera poses, and pose-free alternatives lag behind pose-required methods due to compounding pose-estimation errors and scale ambiguity.
Approach
Predict Gaussians for all input views directly in the canonical coordinate frame of the first view (no local-to-global transform), resolve scale ambiguity by feeding camera intrinsics as a token concatenated with image tokens, and train with photometric loss only.
Key results
- On RealEstate10K average over overlaps, NoPoSplat reaches 25.033 PSNR / 0.838 SSIM, above pose-required MVSplat (23.977) and pixelSplat (23.848), with the largest gain at small overlap (22.514 vs 20.353 PSNR for MVSplat).
- Relative pose estimation on RE10K AUC@5 improves from 0.672 (Ours) to 0.691 (Ours*, trained with DL3DV), exceeding RoMa (0.546) and MASt3R (0.372).
- Feed-forward inference predicts Gaussians from two 256x256 views in 0.015 s (66 fps) on an RTX 4090, about 5x faster than pixelSplat and 2x faster than MVSplat.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Predict Gaussians in canonical space instead of using transform-then-fuse (ablation) | NoPoSplat · RealEstate10K average | PSNR: 24.632 -> 25.033 | Helped |
| Add camera intrinsic token embedding (ablation) | NoPoSplat · RealEstate10K average | PSNR: 23.543 -> 25.033 | Helped |
| Use intrinsic token embedding instead of dense intrinsic embedding (ablation) | NoPoSplat · RealEstate10K average | PSNR: 24.809 -> 25.033 | Helped |
| Use intrinsic token embedding instead of global additive embedding (ablation) | NoPoSplat · RealEstate10K average | PSNR: 24.952 -> 25.033 | Helped |
| Add the RGB image shortcut to the Gaussian parameter prediction head (ablation) | NoPoSplat · RealEstate10K average | PSNR: 24.279 -> 25.033 | Helped |
| Add a third input view (ablation) | NoPoSplat · RealEstate10K | PSNR: 25.033 -> 26.619 | Helped |
| Initialize the backbone with MASt3R rather than DUSt3R weights (ablation) | NoPoSplat · RealEstate10K average | PSNR: 24.553 -> 25.033 | Helped |
| Initialize the backbone with MASt3R rather than CroCoV2 weights (ablation) | NoPoSplat · RealEstate10K average | PSNR: 24.559 -> 25.033 | Helped |
| Add photometric pose refinement after PnP-RANSAC initialization (ablation) | NoPoSplat pose estimation · ScanNet-1500 | AUC 5°: 0.287 -> 0.318 | Helped |
| Add PnP-RANSAC initialization before photometric refinement (ablation) | NoPoSplat pose estimation · ScanNet-1500 | AUC 5°: 0.017 -> 0.318 | Helped |
| Train with RE10K+DL3DV instead of RE10K alone | NoPoSplat pose estimation · RE10K | AUC 5°: 0.672 -> 0.691 | Helped |
| Use NoPoSplat rather than MVSplat | Novel view synthesis · RealEstate10K average | PSNR: 23.977 -> 25.033 | Helped |
Surprising findings
- A pose-free model trained only with photometric loss outperforms pose-required methods such as MVSplat and pixelSplat in novel view synthesis, especially at small overlap.
- Removing intrinsic embedding entirely causes large degradation (PSNR 25.033 to 23.543), indicating scale ambiguity is driven by focal length.
Limitations
- Assumes known camera intrinsics; heuristic focal length used for in-the-wild data.
- Non-generative feed-forward model cannot reconstruct unseen regions with detailed geometry and texture.
- Limited training data (RE10K, ACID, DL3DV) restricts generalization to diverse in-the-wild scenes; only static scenes handled.
Experimental setup
Datasets: RealEstate10K ACID DL3DV DTU ScanNet ScanNet++ ScanNet-1500
Benchmarks: RealEstate10K ACID DTU ScanNet++ ScanNet-1500 pixelSplat evaluation set
Metrics: PSNR SSIM LPIPS AUC@5 AUC@10 AUC@20
Baselines: pixelNeRF AttnRend pixelSplat MVSplat DUSt3R MASt3R Splatt3R CoPoNeRF RoMa
Builds on: pixelSplat MVSplat MASt3R DUSt3R DPT
Closest papers by meaning
- SelfSplat: Pose-Free and 3D Prior-Free Generalizable 3D Gaussian Splatting (CVPR 2025 · similarity 0.93)
- PF3plat: Pose-Free Feed-Forward 3D Gaussian Splatting for Novel View Synthesis (ICML 2025 · similarity 0.92)
- Sparse-view Pose Estimation and Reconstruction via Analysis by Generative Synthesis (NeurIPS 2024 · similarity 0.91)
- FreeSplat: Generalizable 3D Gaussian Splatting Towards Free View Synthesis of Indoor Scenes (NeurIPS 2024 · similarity 0.91)
- COLMAP-Free 3D Gaussian Splatting (CVPR 2024 · similarity 0.90)
- KeyGS: A Keyframe-Centric Gaussian Splatting Method for Monocular Image Sequences (AAAI 2025 · similarity 0.90)
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.