Research map / 3D vision / Face and shape
VGG-T$^3$: Offline Feed-Forward 3D Reconstruction at Scale
Sven Elflein, Ruilong Li, Sérgio Agostinho, Zan Gojcic, Laura Leal-Taixé, Qunjie Zhou, Aljosa Osep
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 |
|---|---|---|---|
| Initialize from pre-trained softmax-attention weights instead of training TTT from scratch (ablation) | VGG-T3 · ScanNet++ 224x224 ablation | mAA(30): 52.95 -> 72.16 | Improved |
| Use test-time-training linearization instead of T2R (ablation) | VGG-T3 · ScanNet++ 224x224 ablation | mAA(30): 66.27 -> 72.16 | Improved |
| Use test-time-training linearization instead of LoLCATs (ablation) | VGG-T3 · ScanNet++ 224x224 ablation | mAA(30): 62.87 -> 72.16 | Improved |
| Add ShortConv2D spatial mixing on values (ablation) | VGG-T3 · ScanNet++ 224x224 ablation | mAA(30): 72.16 -> 74.14 | Improved |
| Add ShortConv2D spatial mixing on values (ablation) | VGG-T3 · ScanNet++ 224x224 ablation | CD: 0.074 -> 0.066 | Improved |
| Use 2 TTT optimizer steps instead of 1 at test time for large collections (ablation) | VGG-T3 · Chamfer distance for up to 2000 views | Chamfer Distance | Improved |
| Further increase optimizer steps beyond 2 (ablation) | VGG-T3 · large image collections | reconstruction quality | No clear effect |
| Replace LayerNorm with L2 normalization on queries and keys | VGG-T3 linearization · convergence from pre-trained weights | convergence speed | Improved |
| Replace softmax global attention with TTT MLP (linear time) | VGGT · 1k images on 7scenes | runtime: VGGT over 11 minutes -> 58 seconds (11.6x) | Improved |
| Replace softmax global attention with TTT MLP (linear time) | VGGT · ScanNet, Sintel, TUM camera pose | ATE: 0.035 -> 0.070 (ScanNet) | Hurt |
| Replace softmax global attention with TTT MLP (linear time) | VGGT · pointmap estimation DTU | Chamfer Distance: 1.537 -> 1.654 | Mixed |
| Shard images across GPUs for distributed TTT inference | VGG-T3 · 2000 images | latency (s): 230.7 -> 48.5 (1 to 4 GPUs) | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- TTT3R — TTT3R: 3D Reconstruction as Test-Time Training (ICLR 2026) · better on 10 of 12 shared results
- VGGT — VGGT: Visual Geometry Grounded Transformer (CVPR 2025) · better on 1 of 10 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
- VGGT: Visual Geometry Grounded Transformer (CVPR 2025 · similarity 0.92)
- VGGT-Ω (CVPR 2026 · similarity 0.92)
- TurboVGGT: Fast Visual Geometry Reconstruction with Adaptive Alternating Attention (NeurIPS 2026 · similarity 0.89)
- MVSplat360: Feed-Forward 360 Scene Synthesis from Sparse Views (NeurIPS 2024 · similarity 0.88)
- Fin3R: Fine-tuning Feed-forward 3D Reconstruction Models via Monocular Knowledge Distillation (NeurIPS 2025 · similarity 0.88)
- VLM-3R: Vision-Language Models Augmented with Instruction-Aligned 3D Reconstruction (CVPR 2026 · 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.
