Research map / 3D vision / Human pose estimation
True Self-Supervised Novel View Synthesis is Transferable
Thomas Mitchel, Hyunwoo Ryu, Vincent Sitzmann
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add transferability objective with masked augmentation vs plain autoencoding (unconstrained) (ablation) | stereo-monocular XFactor · RE10K transfer test | TPS AUC@20: 36.8 -> 47.2 | Improved |
| add transferability objective vs 16-dim bottleneck (ablation) | stereo-monocular XFactor · RE10K transfer test | TPS AUC@20: 40.6 -> 47.2 | Improved |
| add transferability objective vs 16-dim bottleneck (ablation) | stereo-monocular XFactor · RE10K pose probe | probe AUC@20: 47.2 -> 54.8 | Improved |
| add SE(3) and Plucker parameterization to the stereo-monocular model (ablation) | unconstrained stereo-monocular model · DL3DV transfer test | TPS AUC@20: 39.4 -> 26.4 | Hurt |
| add an extra context view to the renderer (ablation) | stereo-monocular XFactor · DL3DV transfer test | TPS AUC@20: 44.8 -> 25.1 | Hurt |
| add an extra context view to both pose encoder and renderer (ablation) | stereo-monocular XFactor · MVImgNet transfer test | TPS AUC@20: 49.3 -> 1.2 | Hurt |
| add temporal-shift corruption (+1) to training pairs (ablation) | stereo-monocular XFactor · DL3DV transfer test | TPS AUC@20: 44.8 -> 36.2 | Hurt |
| add shift +1 and +2 corruption to training pairs (ablation) | stereo-monocular XFactor · DL3DV pose probe | probe AUC@20: 61.2 -> 47.0 | Hurt |
| replace pose-preserving augmentation objective with SimCLR (ablation) | pose encoder · RE10K pose probe | probe AUC@20: 54.8 -> 3.8 | Hurt |
| replace transferability objective with VICReg (ablation) | pose encoder · RE10K pose probe | probe AUC@20: 54.8 -> 0.8 | Hurt |
| XFactor vs RayZer on transferability test | XFactor multi-view · RE10K transfer test | TPS AUC@20: 7.6 -> 55.2 | Improved |
| XFactor vs RUST on transferability test | XFactor multi-view · CO3Dv2 transfer test | TPS AUC@20: 5.4 -> 31.2 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- RayZer — RayZer: A Self-supervised Large View Synthesis Model (ICCV 2025) · better on 8 of 8 shared results
- RUST — RUST: Latent Neural Scene Representations From Unposed Imagery (CVPR 2023) · better on 8 of 8 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Less Decoder is More Encoder: Geometric Representation Learning from Novel View Synthesis (NeurIPS 2026 · similarity 0.88)
- MAPConNet: Self-supervised 3D Pose Transfer with Mesh and Point Contrastive Learning (ICCV 2023 · similarity 0.87)
- Novel View Synthesis with View-Dependent Effects from a Single Image (CVPR 2024 · similarity 0.86)
- Self-Supervised Cross-View Correspondence with Predictive Cycle Consistency (CVPR 2025 · similarity 0.86)
- Novel View Synthesis with Pixel-Space Diffusion Models (CVPR 2025 · similarity 0.86)
- Unifying Correspondence Pose and NeRF for Generalized Pose-Free Novel View Synthesis (CVPR 2024 · similarity 0.86)
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.
