Research map / Label-efficient and robust learning / Self-supervised and contrastive
A Closer Look at Invariances in Self-Supervised Pre-training for 3D Vision
Lanxiao Li, Michael Heizmann
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 |
|---|---|---|---|
| pre-train with DPCo (depth-point contrast) instead of training from scratch | VoteNet (PointNet++) · ScanNet detection | AP50: 37.6 -> 41.5 | Improved |
| contrast a 2D format (depth map) with point clouds instead of point-vs-voxel (3D-3D) (ablation) | VoteNet (PointNet++) · ScanNet detection | AP25: 62.8 -> 64.2 | Improved |
| add voxel-point format contrast (PVCo) over augmentation-only PPCo (ablation) | VoteNet (PointNet++) · SUN RGB-D detection | AP25: 58.6 -> 59.3 | No clear effect |
| use explicit perspective-invariance (PointContrast) instead of augmentation-only PPCo (ablation) | VoteNet (PointNet++) · SUN RGB-D and ScanNet detection | AP50: 34.9 -> 34.1 (SUN RGB-D), 39.5 -> 38.1 (ScanNet) | Mixed |
| replace RGB image modality (IPCo) with depth map format (DPCo) (ablation) | VoteNet (PointNet++) · SUN RGB-D detection | AP50: 35.5 -> 35.6 | No clear effect |
| add local correspondence only (DPCo) vs from scratch (ablation) | VoteNet (PointNet++) · ScanNet detection | AP25: 60.0 -> 64.2 | Improved |
| add global correspondence only (DPCo) vs from scratch (ablation) | VoteNet (PointNet++) · ScanNet detection | AP25: 60.0 -> 63.8 | Improved |
| combine local and global correspondence (DPCo) instead of local only (ablation) | VoteNet (PointNet++) · SUN RGB-D detection | AP50: 35.1 -> 35.6 | Improved |
| DPCo pre-training with only 5% of ScanNet training data | VoteNet (PointNet++) · ScanNet detection (5% data) | AP25: 13.3 -> 36.5 | Improved |
| DPCo pre-training with only 5% of ScanNet training data | VoteNet (PointNet++) · ScanNet detection (5% data) | AP50: 2.4 -> 14.4 | Improved |
| pre-train depth map encoder with depth-depth contrast (DDCo) vs from scratch | 2.5D-VoteNet · SUN RGB-D detection | AP25: 60.8 -> 56.0 | Hurt |
| add point cloud contrast (DPCo) to depth map encoder pre-training instead of depth-depth only (ablation) | 2.5D-VoteNet · SUN RGB-D detection | AP25: 56.0 -> 61.4 | Improved |
Closest papers by meaning
- Demystifying Contrastive Self-Supervised Learning: Invariances, Augmentations and Dataset Biases (NeurIPS 2020 · similarity 0.89)
- Transitive Invariance for Self-Supervised Visual Representation Learning (ICCV 2017 · similarity 0.89)
- Invaria: Learning Scale and Density Invariance in Point Clouds via Next-Resolution Prediction (NeurIPS 2026 · similarity 0.86)
- Building a Strong Pre-Training Baseline for Universal 3D Large-Scale Perception (CVPR 2024 · similarity 0.86)
- In-Context Symmetries: Self-Supervised Learning through Contextual World Models (NeurIPS 2024 · similarity 0.86)
- Harnessing small projectors and multiple views for efficient vision pretraining (NeurIPS 2024 · 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.
