Research map / 3D vision / Gaussian splatting
Epipolar-Free 3D Gaussian Splatting for Generalizable Novel View Synthesis
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 eFreeSplat, a feed-forward generalizable 3DGS model that avoids epipolar priors by using a CroCo v2-pretrained ViT-B with cross-attention decoder and an iterative cross-view Gaussians alignment module, evaluated on RealEstate10K and ACID.
Problem
Generalizable feed-forward 3D Gaussian splatting from sparse views relies on epipolar priors, which are unreliable in non-overlapping and occluded regions of complex real-world scenes.
Approach
Replace epipolar line sampling with cross-view completion pre-trained ViT and cross-attention decoder for epipolar-free feature matching, plus an iterative alignment module that uses homography-warped features to correct inconsistent per-view depth scales before merging Gaussians.
Key results
- On RealEstate10K, eFreeSplat reaches 26.45 PSNR / 0.865 SSIM / 0.126 LPIPS vs. MVSplat 26.39 / 0.869 / 0.128.
- On ACID, eFreeSplat reaches 28.30 PSNR / 0.851 SSIM / 0.140 LPIPS, versus MVSplat 28.25 / 0.843 / 0.144.
- In the low-overlap subsets (overlap below 0.5), eFreeSplat reports 25.46 PSNR vs. 24.64 for MVSplat.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use eFreeSplat instead of MVSplat | Generalizable novel view synthesis · RealEstate10K | PSNR: 26.39 -> 26.45 | Helped |
| use eFreeSplat instead of MVSplat | Generalizable novel view synthesis · ACID | PSNR: 28.25 -> 28.30 | Helped |
| use eFreeSplat instead of MVSplat | Generalizable novel view synthesis · RealEstate10K | SSIM: 0.869 -> 0.865 | Hurt |
| use eFreeSplat instead of MVSplat | Generalizable novel view synthesis · RealEstate10K | LPIPS: 0.128 -> 0.126 | Helped |
| add epipolar-free cross-view mutual perception (ablation) | eFreeSplat · RealEstate10K | PSNR: 22.04 -> 26.45 | Helped |
| add iterative cross-view Gaussians alignment (ablation) | eFreeSplat · RealEstate10K | PSNR: 23.03 -> 26.45 | Helped |
| add self-supervised cross-view completion pre-training (ablation) | eFreeSplat · RealEstate10K | PSNR: 24.81 -> 26.45 | Helped |
| increase alignment iterations from 1 to 2 (ablation) | eFreeSplat · RealEstate10K | PSNR: 23.36 -> 26.45 | Helped |
| increase alignment iterations from 2 to 3 (ablation) | eFreeSplat · RealEstate10K | PSNR: 26.45 -> 26.40 | No clear effect |
| fine-tune CroCo pre-trained weights on RealEstate10K (ablation) | eFreeSplat · RealEstate10K test set | PSNR: 26.45 -> 26.32 | Hurt |
| train CroCo on RealEstate10K without loading pre-trained weights (ablation) | eFreeSplat · RealEstate10K | PSNR: 24.81 -> 25.12 | Helped |
Surprising findings
- Removing CroCo pre-training drops PSNR to 24.81 from 26.45, and the model is reported to fail when input views are very close, attributed to a distribution gap between CroCo pretraining data (overlap at most 0.75) and RealEstate10K/ACID (overlap 0.9-1.0).
Limitations
- Model lacks geometric inductive biases, making it data-hungry and sensitive to training data distribution.
- Per-pixel Gaussian mapping struggles to reconstruct occluded or unseen parts of the scene.
- Unstable results when input viewpoints are very close due to CroCo pretraining distribution mismatch.
Experimental setup
Datasets: RealEstate10K ACID large-scale cross-view datasets for CroCo v2 pretraining
Benchmarks: RealEstate10K ACID
Metrics: PSNR SSIM LPIPS inference time
Baselines: Du et al. GPNR pixelSplat MVSplat
Builds on: 3D Gaussian Splatting CroCo v2 pixelSplat MVSplat
Closest papers by meaning
- FreeSplat: Generalizable 3D Gaussian Splatting Towards Free View Synthesis of Indoor Scenes (NeurIPS 2024 · similarity 0.90)
- PF3plat: Pose-Free Feed-Forward 3D Gaussian Splatting for Novel View Synthesis (ICML 2025 · similarity 0.89)
- TranSplat: Generalizable 3D Gaussian Splatting from Sparse Multi-View Images with Transformers (AAAI 2025 · similarity 0.89)
- No Pose, No Problem: Surprisingly Simple 3D Gaussian Splats from Sparse Unposed Images (ICLR 2025 · similarity 0.89)
- NexusGS: Sparse View Synthesis with Epipolar Depth Priors in 3D Gaussian Splatting (CVPR 2025 · similarity 0.88)
- SelfSplat: Pose-Free and 3D Prior-Free Generalizable 3D Gaussian Splatting (CVPR 2025 · similarity 0.88)
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.