Research map / 3D vision / 3D generation
Pri3D: Can 3D Priors Help 2D Representation Learning?
Ji Hou, Saining Xie, Benjamin Graham, Angela Dai, Matthias Nießner
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 Pri3D pre-training (view+geo) on ScanNet after ImageNet init | ResNet50 U-Net · ScanNet semantic segmentation | mIoU: 55.7 -> 61.7 (+6.0) | Improved |
| add view-invariant contrastive loss (ablation) | Pri3D ResNet50 · ScanNet semantic segmentation | mIoU: 55.7 -> 61.3 | Improved |
| add geometric prior contrastive loss (together with view loss) (ablation) | Pri3D ResNet18 · ScanNet semantic segmentation | mIoU: 54.4 -> 55.7 | Improved |
| add geometric prior to view loss (ablation) | Pri3D ResNet50 · NYUv2 semantic segmentation | mIoU: 54.2 -> 54.7 | No clear effect |
| add ScanNet data to MoCoV2 without 3D priors (combine) | MoCoV2 ImageNet · ScanNet semantic segmentation | mIoU: 55.7 -> 54.9 | Hurt |
| add ScanNet data to MoCoV2 by mixing batches | MoCoV2 ImageNet · ScanNet semantic segmentation | mIoU: 55.7 -> 54.5 | Hurt |
| replace trivial color-augmentation correspondences with multi-view 3D correspondences (ablation) | Pri3D ResNet50 · ScanNet semantic segmentation | mIoU: 56.4 -> 61.3 | Improved |
| replace single-frame depth prediction pre-training with Pri3D view+geo (ablation) | ResNet50 · ScanNet semantic segmentation | mIoU: 58.4 -> 61.7 | Improved |
| add depth proxy loss (ablation) | Pri3D (View) · ScanNet semantic segmentation | mIoU: 60.2 -> 61.3 | Improved |
| add Pri3D to DeepLabV3 | DeepLabV3 ResNet50 · ScanNet semantic segmentation | mIoU: 57.0 -> 61.3 | Improved |
| add Pri3D to PSPNet | PSPNet ResNet50 · ScanNet semantic segmentation | mIoU: 59.7 -> 62.8 | Improved |
| add Pri3D to ImageNet init on NYUv2 detection | ResNet50 · NYUv2 object detection | AP@0.5: 29.9 -> 34.0 | Improved |
Closest papers by meaning
- Mask3D: Pre-Training 2D Vision Transformers by Learning Masked 3D Priors (CVPR 2023 · similarity 0.91)
- PonderV2: Improved 3D Representation With a Universal Pre-Training Paradigm (TPAMI 2025 · similarity 0.90)
- ImageNet3D: Towards General-Purpose Object-Level 3D Understanding (NeurIPS 2024 · similarity 0.88)
- Building a Strong Pre-Training Baseline for Universal 3D Large-Scale Perception (CVPR 2024 · similarity 0.87)
- Multi-View Representation is What You Need for Point-Cloud Pre-Training (ICLR 2024 · similarity 0.86)
- Beyond 3D VQAs: Injecting 3D Spatial Priors into Vision-Language Models for Enhanced Geometric Reasoning (CVPR 2026 · 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.
