Research map / Detection and segmentation / 3D detection and driving
4DContrast: Contrastive Learning with Dynamic Correspondences for 3D Scene Understanding
Yujin Chen, Matthias Nießner, Angela Dai
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 4D scene-object augmentation to 3D-only pre-training (ablation) | 4DContrast · ScanNet semantic segmentation (fine-tuning) | mIoU: 71.5 -> 71.7 | Improved |
| add 3D-4D and 4D-4D contrastive constraints (ablation) | 4DContrast · ScanNet semantic segmentation (fine-tuning) | mIoU: 71.7 -> 72.3 | Improved |
| add 4D priors on top of 3D-only variant (ablation) | 4DContrast · ScanNet instance segmentation | mAP@0.5: 54.1 -> 57.6 | Improved |
| add 4D priors on top of 3D-only variant (ablation) | 4DContrast · ScanNet object detection | mAP@0.5: 38.7 -> 40.0 | Improved |
| add 4D priors on top of 3D-only variant (1% data) (ablation) | 4DContrast · ScanNet semantic segmentation (1% data) | mIoU: 27.0 -> 28.2 | Improved |
| add 4D priors on top of 3D-only variant (100% data) (ablation) | 4DContrast · ScanNet semantic segmentation (100% data) | mIoU: 71.5 -> 72.3 | Improved |
| use 4D backbone instead of two 3D backbones (ablation) | 4DContrast · ScanNet semantic segmentation (fine-tuning) | mIoU: 71.8 -> 72.3 | Improved |
| use SimSiam instead of SimCLR (ablation) | 4DContrast · ScanNet semantic segmentation (fine-tuning) | mIoU: 71.6 -> 72.3 | Improved |
| use sequence length 4 instead of 5 (ablation) | 4DContrast · ScanNet semantic segmentation (fine-tuning) | mIoU: 71.0 -> 72.3 | Improved |
| use sequence length 4 instead of 3 (ablation) | 4DContrast · ScanNet semantic segmentation (fine-tuning) | mIoU: 71.9 -> 72.3 | Improved |
| 4DContrast pre-training vs training from scratch | VoteNet · ScanNet object detection | mAP@0.5: 34.5 -> 40.0 | Improved |
| 4DContrast pre-training vs training from scratch | VoteNet · SUN RGB-D object detection | mAP@0.5: 31.7 -> 38.2 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- H3DNet — H3DNet: 3D Object Detection Using Hybrid Geometric Primitives (ECCV 2020) · better on 1 of 1 shared result
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
- Motion4D: Learning 3D-Consistent Motion and Semantics for 4D Scene Understanding (NeurIPS 2025 · similarity 0.89)
- Vision-Language Pre-training with Object Contrastive Learning for 3D Scene Understanding (AAAI 2024 · similarity 0.88)
- Building a Strong Pre-Training Baseline for Universal 3D Large-Scale Perception (CVPR 2024 · similarity 0.88)
- ImageNet3D: Towards General-Purpose Object-Level 3D Understanding (NeurIPS 2024 · similarity 0.87)
- Compositional 4D Dynamic Scenes Understanding with Physics Priors for Video Question Answering (ICLR 2025 · similarity 0.87)
- Uni4D: Unifying Visual Foundation Models for 4D Modeling from a Single Video (CVPR 2025 · similarity 0.87)
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.
