Research map / 3D vision / Human pose estimation
FLARE: Feed-forward Geometry, Appearance and Camera Estimation from Uncalibrated Sparse Views
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.
FLARE is a feed-forward transformer that cascades camera pose estimation, two-stage point-map geometry, and 3D Gaussian appearance to recover poses, geometry, and novel views from uncalibrated 2-8 sparse images in under 0.5 s.
Problem
Recovering accurate camera poses, 3D geometry and appearance from uncalibrated sparse-view images, where SfM/MVS fail due to limited overlap and optimization-based or two-view feed-forward methods are slow or inaccurate.
Approach
Uses estimated camera poses as proxies in a cascaded pipeline: a neural pose predictor initializes poses, a transformer predicts camera-centric point maps that a learned global geometry projector maps to world coordinates, and a Gaussian head renders novel views, all trained jointly end-to-end.
Key results
- On RealEstate10K pose estimation (5 views), FLARE reaches AUC@30 of 79.9 versus 61.1 for MASt3R and 51.6 for PoseDiffusion.
- On RealEstate10K 2-view novel view synthesis, FLARE achieves PSNR 23.765 and LPIPS 0.191, outperforming pose-required MVSplat (22.568 PSNR) without camera poses as input.
- Inference takes under 0.5 seconds for feed-forward reconstruction from 2-8 uncalibrated views; training took about 14 days on 64 A800 GPUs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use camera-pose estimates as geometry-learning conditioning (ablation) | FLARE · BlendedMVS test set | Overall Chamfer distance: 0.0319 -> 0.0288 | Helped |
| Predict camera-centric point maps before global geometry instead of directly predicting global geometry (ablation) | FLARE · BlendedMVS test set | Overall Chamfer distance: 0.0295 -> 0.0288 | Helped |
| Jointly train the pose predictor and reconstructor (ablation) | FLARE · BlendedMVS test set | Overall Chamfer distance: 0.0299 -> 0.0288 | Helped |
| Use a DPT decoder rather than a shallow MLP for point-map regression (ablation) | FLARE · BlendedMVS test set | Overall Chamfer distance: 0.0436 -> 0.0288 | Helped |
| Add rendering loss during training (ablation) | FLARE · BlendedMVS test set | ACC.: 0.0250 -> 0.0320 | Hurt |
| Add rendering loss during training (ablation) | FLARE · BlendedMVS test set | COMP.: 0.0325 -> 0.0244 | Helped |
Surprising findings
- The pose-free FLARE outperforms pose-required baselines MVSplat and pixelSplat on RealEstate10K and DL3DV novel view synthesis.
- Adding the rendering loss improves completeness but slightly worsens accuracy in the geometry ablation.
Limitations
- Trained with 8 views; generalization beyond that is evaluated but depends on memory limits.
- Rendering loss trades off geometry accuracy against completeness.
- Pose augmentation is needed because inaccurate coarse poses at inference harm the pipeline.
Experimental setup
Datasets: MegaDepth ARKitScenes BlendedMVS ScanNet++ CO3D-v2 Waymo WildRGBD DL3DV RealEstate10K DTU ETH3D TUM RGBD
Benchmarks: RealEstate10K DTU ETH3D TUM RGBD DL3DV BlendedMVS
Metrics: RRA@5 RTA@5 AUC@30 accuracy completeness Chamfer overall PSNR SSIM LPIPS
Baselines: DUSt3R MASt3R Spann3R COLMAP PixSfM VGGSfM RelPose PoseDiffusion CoPoNeRF Splatt3R pixelSplat MVSplat
Builds on: DUSt3R MASt3R VGGSfM PoseDiffusion pixelSplat
Closest papers by meaning
- Sparse-view Pose Estimation and Reconstruction via Analysis by Generative Synthesis (NeurIPS 2024 · similarity 0.88)
- LEAP: Liberate Sparse-View 3D Modeling from Camera Poses (ICLR 2024 · similarity 0.86)
- FlareX: A Physics-Informed Dataset for Lens Flare Removal via 2D Synthesis and 3D Rendering (NeurIPS 2025 · similarity 0.86)
- PF3plat: Pose-Free Feed-Forward 3D Gaussian Splatting for Novel View Synthesis (ICML 2025 · similarity 0.85)
- Improving Robustness for Joint Optimization of Camera Pose and Decomposed Low-Rank Tensorial Radiance Fields (AAAI 2024 · similarity 0.85)
- MS-GS: Multi-Appearance Sparse-View 3D Gaussian Splatting in the Wild (NeurIPS 2025 · similarity 0.85)
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.