Research map / Label-efficient and robust learning / Self-supervised and contrastive
Contrast with Reconstruct: Contrastive 3D Representation Learning Guided by Generative Pretraining
Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, Li Yi
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 |
|---|---|---|---|
| Guide cross-modal contrastive learning with masked reconstruction | ReCon · ScanObjectNN PB_T50_RS | accuracy: 90.63% vs. 82.48% | Improved |
| Add reconstruction guidance to cross-modal contrastive pretraining (ablation) | ReCon · ScanObjectNN | accuracy: 88.42 -> 90.18 | Improved |
| Add image contrastive supervision alongside reconstruction and text supervision (ablation) | ReCon · ScanObjectNN | accuracy: 89.77 -> 90.63 | Improved |
| Add self-contrastive supervision to reconstruction, text, and image objectives (ablation) | ReCon · ScanObjectNN | accuracy: 90.63 -> 89.98 | Hurt |
| Add stop-gradient to ReCon-block cross-attention (ablation) | ReCon · ScanObjectNN | accuracy: 81.60 -> 90.63 | Improved |
| Add stop-gradient to ReCon-block cross-attention (ablation) | ReCon · ModelNet40 | accuracy: 89.7 -> 94.1 | Improved |
| Replace Smooth ℓ1 with InfoNCE (ablation) | ReCon contrastive objective · ScanObjectNN | accuracy: 90.63 -> 90.11 | Hurt |
| Replace Smooth ℓ1 with InfoNCE (ablation) | ReCon contrastive objective · ModelNet40 | accuracy: 94.1 -> 93.8 | Hurt |
| Use paired rather than category-shuffled image data during pretraining (ablation) | ReCon · ScanObjectNN | accuracy: 90.22 -> 90.63 | Improved |
| Use paired rather than category-shuffled image data during pretraining (ablation) | ReCon · ModelNet40 zero-shot | accuracy: 60.4 -> 66.8 | Improved |
| Ensemble multiple text prompts (ablation) | ReCon zero-shot classifier · ModelNet40 | top-1 accuracy: 60.6 -> 61.7 | Improved |
| Use Rotation rather than no augmentation (ablation) | ReCon fine-tuning · ScanObjectNN | accuracy: 87.44 -> 90.63 | Improved |
Closest papers by meaning
- The Power of Contrast for Feature Learning: A Theoretical Analysis (JMLR 2023 · similarity 0.87)
- ReconViaGen: Towards Accurate Multi-view 3D Object Reconstruction via Generation (ICLR 2026 · similarity 0.87)
- Decompositional Neural Scene Reconstruction with Generative Diffusion Prior (CVPR 2025 · similarity 0.86)
- Do Generated Data Always Help Contrastive Learning? (ICLR 2024 · similarity 0.86)
- $\mathbb{X}$-Sample Contrastive Loss: Improving Contrastive Learning with Sample Similarity Graphs (ICLR 2025 · similarity 0.86)
- On the Similarities of Embeddings in Contrastive Learning (ICML 2025 · 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.
