Research map / 3D vision / 3D generation
4D Contrastive Superflows are Dense 3D Representation Learners
Xiang Xu, Lingdong Kong, Hui Shuai, Wenwei Zhang, Liang Pan, Kai Chen, Ziwei Liu, Qingshan Liu
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 view consistency alignment of superpixels (ablation) | SuperFlow · nuScenes linear probing | mIoU: 44.65 -> 45.57 | Improved |
| add dense-to-sparse consistency regularization (ablation) | SuperFlow · nuScenes linear probing | mIoU: 45.57 -> 46.17 | Improved |
| add flow-based contrastive learning (ablation) | SuperFlow · nuScenes linear probing | mIoU: 45.57 -> 47.24 | Improved |
| use three frames instead of two for temporal contrast (ablation) | SuperFlow · nuScenes linear probing | mIoU: 46.39 -> 47.66 | Improved |
| lengthen the timespan between frames beyond 0.5s (ablation) | SuperFlow · nuScenes linear probing | mIoU: 47.66 -> 46.20 (2.0s) | Hurt |
| concatenate more sweeps (7) for dense point cloud (ablation) | SuperFlow · nuScenes linear probing | mIoU: 47.66 -> 46.03 | Hurt |
| use 1 sweep instead of 2 (ablation) | SuperFlow · nuScenes linear probing | mIoU: 47.66 -> 47.41 | Hurt |
| scale 3D backbone from MinkUNet-34 to MinkUNet-50 (ablation) | SuperFlow · nuScenes linear probing | mIoU: 47.66 -> 54.11 | Improved |
| scale 3D backbone to MinkUNet-101 (ablation) | SuperFlow · nuScenes linear probing | mIoU: 54.11 -> 52.56 | Hurt |
| scale the 2D distillation backbone from ViT-S to ViT-L | SuperFlow · nuScenes linear probing | mIoU: 46.44 -> 48.01 | Improved |
| SuperFlow vs Seal on nuScenes-C linear-probe mRR | SuperFlow · nuScenes-C linear probing | mRR: 75.52 vs 78.15 (PPKT) | Mixed |
Closest papers by meaning
- Enhanced Spatiotemporal Consistency for Image-to-LiDAR Data Pretraining (TPAMI 2025 · similarity 0.96)
- TREND: Unsupervised 3D Representation Learning via Temporal Forecasting for LiDAR Perception (NeurIPS 2025 · similarity 0.87)
- ICP-Flow: LiDAR Scene Flow Estimation with ICP (CVPR 2024 · similarity 0.87)
- Fine-grained Image-to-LiDAR Contrastive Distillation with Visual Foundation Models (NeurIPS 2024 · similarity 0.86)
- CALICO: Self-Supervised Camera-LiDAR Contrastive Pre-training for BEV Perception (ICLR 2024 · similarity 0.86)
- LSK3DNet: Towards Effective and Efficient 3D Perception with Large Sparse Kernels (CVPR 2024 · 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.
