Research map / 3D vision / 3D generation
Learning Viewpoint-Agnostic Visual Representations by Recovering Tokens in 3D Space
Jinghuan Shang, Srijan Das, Michael S Ryoo
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 3DTRL | Swin-T · Pouring video alignment | Kendall's tau: 0.584 -> 0.683 | Improved |
| add 3DTRL | Swin-T · CIFAR-100 | accuracy: 21.53 -> 21.55 | No clear effect |
| add 3DTRL | TnT-S · CIFAR-100 | accuracy: 54.07 -> 56.00 | Improved |
| add 3DTRL | TimeSformer · Smarthome CV2 mPA | mPA: 27.5 -> 34.0 | Improved |
| pre-train 3DTRL on K400 versus random init | TimeSformer + 3DTRL · Smarthome CV2 | accuracy: 69.5 -> 71.9 | Improved |
| use joint-temporal (single camera) instead of per-frame camera estimation (ablation) | TimeSformer + 3DTRL · Smarthome CV2 | accuracy: 62.9 -> 58.6 | Hurt |
| use joint-temporal camera estimation with K400 pre-training (ablation) | TimeSformer + 3DTRL · Smarthome CV2 | accuracy: 71.9 -> 68.2 | Hurt |
| add perspective augmentation instead of 3DTRL (ablation) | DeiT · Pick video alignment | Kendall's tau: 0.244 -> -0.249 | Hurt |
| add 3DTRL | DeiT · Pick video alignment | Kendall's tau: 0.244 -> 0.635 | Improved |
| use multiple 3DTRL modules (layers 4, 6, 8) (ablation) | DeiT-T + 3DTRL · CIFAR-10 | accuracy: 78.8 -> 79.5 | Improved |
| disable Mixup and CutMix (ablation) | DeiT-T · ImageNet-1K-Perturbed | accuracy: 61.3 -> 54.5 | Hurt |
| disable Mixup and CutMix (ablation) | DeiT-T + 3DTRL · ImageNet-1K-Perturbed | accuracy: 64.6 -> 63.0 | Hurt |
Closest papers by meaning
- ImageNet3D: Towards General-Purpose Object-Level 3D Understanding (NeurIPS 2024 · similarity 0.86)
- Cameras as Relative Positional Encoding (NeurIPS 2025 · similarity 0.86)
- Imagine3D-LLM: Teaching MLLMs to Imagine 3D Scenes Before Answering (NeurIPS 2026 · similarity 0.85)
- VLM-3R: Vision-Language Models Augmented with Instruction-Aligned 3D Reconstruction (CVPR 2026 · similarity 0.85)
- Understanding Multi-View Transformers (NeurIPS 2026 · similarity 0.85)
- Learning from Videos for 3D World: Enhancing MLLMs with 3D Vision Geometry Priors (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.
