Research map / 3D vision / Face and shape
2D-LFM: Lifting Foundation Model without 3D Supervision
Mosam Dabhi, Irhas Gill, László A. Jeni, Simon Lucey
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 |
|---|---|---|---|
| inject positional encoding at every attention layer instead of only at the input (ablation) | 2D-only transformer lifter · Pascal3D+ | MPJPE (mm): 100.3 -> 8.1 | Improved |
| inject positional encoding at every attention layer instead of only at the input (ablation) | 2D-only transformer lifter · Human3.6M | MPJPE (mm): 63.4 -> 33.1 | Improved |
| inject PE only at first and last layers (ablation) | 2D-only transformer lifter · Pascal3D+ | MPJPE (mm): 100.3 -> 92.1 | No clear effect |
| inject PE at every 2nd layer rather than every layer (ablation) | 2DLFM · Pascal3D+ | MPJPE (mm): 8.1 -> 26.1 | Hurt |
| use analytical RFF/Fourier PE instead of ViT-style sinusoidal PE (ablation) | 2DLFM · Pascal3D+ | MPJPE (mm): 92.3 -> 8.1 | Improved |
| swap Fourier PE for graph Laplacian PE (both per-layer) (ablation) | 2DLFM · Pascal3D+ and Human3.6M | MPJPE (mm): RFF 11.2/38.1 vs Laplacian 9.3/35.8 | No clear effect |
| increase transformer depth with per-layer PE (ablation) | 2DLFM · Pascal3D+ | MPJPE (mm): 15.3 (4 layers) -> 9.3 (12 layers) -> 8.1 (24 layers) | Improved |
| train a single combined multi-category model instead of per-category models (ablation) | 2DLFM · Pascal3D+ bottle | MPJPE (mm): 100 -> 7.2 | Improved |
| train a single combined multi-category model instead of per-category models (ablation) | 2DLFM · Pascal3D+ car | MPJPE (mm): 8.9 -> 6.7 | Improved |
| learn from 2D supervision only vs 3D-supervised 3D-LFM | 2DLFM · Human3.6M | MPJPE (mm): 46.3 -> 30.9 | Improved |
| learn from 2D supervision only vs 3D-supervised 3D-LFM | 2DLFM · Pascal3D+ | MPJPE (mm): 5.2 -> 8.1 | Hurt |
| use landmark lifting instead of back-projecting landmarks through VGGT depth | 2DLFM · Human3.6M | MPJPE (mm): 107.8 -> 30.9 | Improved |
Closest papers by meaning
- 3D-LFM: Lifting Foundation Model (CVPR 2024 · similarity 0.94)
- Lift3D Policy: Lifting 2D Foundation Models for Robust 3D Robotic Manipulation (CVPR 2025 · similarity 0.88)
- Lift3D: Zero-Shot Lifting of Any 2D Vision Model to 3D (CVPR 2024 · similarity 0.87)
- PandaPose: 3D Human Pose Lifting from a Single Image via Propagating 2D Pose Prior to 3D Anchor Space (NeurIPS 2025 · similarity 0.86)
- Lifting by Image – Leveraging Image Cues for Accurate 3D Human Pose Estimation (AAAI 2024 · similarity 0.86)
- Unsupervised 3D Structure Inference from Category-Specific Image Collections (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.
