Research map / 3D vision / Depth and stereo
ICP-Flow: LiDAR Scene Flow Estimation with ICP
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ICP-Flow, a learning-free LiDAR scene flow estimator that clusters points, uses histogram-voting translation initialization for ICP to estimate per-object rigid transforms, and distills the pseudo labels into a real-time feedforward network.
Problem
LiDAR scene flow estimation for autonomous driving: existing methods predict unconstrained per-point flow that ignores rigid object motion, and either need large-scale training or slow test-time optimization.
Approach
Treat scene flow as per-object rigid transformations: cluster non-ground points, pair clusters over time, and align them with conventional ICP initialized by a histogram vote over point-wise translation vectors, with no learning required for the core estimator.
Key results
- On Waymo, dynamic-foreground EPE of 0.0799 m vs 0.0983 m-level NSFP baseline (about 1.5 cm better per point), with Acc-S improved by over 10%.
- Ours+FNN runs at 21 ms per sample on Waymo, over 1000x faster than NSFP and Chodosh et al. (>1 min), with only marginal performance loss.
- On Waymo over 0.4 s horizons (Table 4), Ours+Tracker improves dynamic-foreground EPE over Ours by about 2 cm and beats the supervised PCA in EPE (0.1799 vs 0.1970).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ICP-Flow instead of NSFP | Waymo scene flow estimation · Waymo, dynamic foreground | EPE (m): 0.0966 -> 0.0799 | Helped |
| Use ICP-Flow pseudo labels instead of NSFP pseudo labels to train the feedforward network | ZeroFlow-style feedforward network · Waymo, dynamic foreground | EPE (m): 0.2229 -> 0.1254 | Helped |
| Replace ICP-Flow with a feedforward network trained on its pseudo labels | ICP-Flow · Waymo, dynamic foreground | EPE (m): 0.0799 -> 0.1254 | Hurt |
| Use ICP-Flow instead of RigidFlow | Cluster-and-ICP scene flow estimation · Waymo, dynamic foreground | EPE (m): 0.2575 -> 0.0799 | Helped |
| Use ICP-Flow pseudo labels instead of NSFP pseudo labels to train the feedforward network | ZeroFlow-style feedforward network · Argoverse-v2, dynamic foreground | EPE (m): 0.2244 -> 0.1697 | Helped |
| Use ICP-Flow instead of the feedforward network trained on its pseudo labels | ICP-Flow · Argoverse-v2, dynamic foreground | EPE (m): 0.1697 -> 0.1653 | Helped |
| Use ICP-Flow instead of PCA | Scene flow estimation · nuScenes, dynamic foreground | EPE (m): 0.1340 -> 0.1445 | Hurt |
| Add cluster association and tracking across intermediate scans | ICP-Flow · Waymo, dynamic foreground, up to 0.4 seconds | EPE (m): 0.2209 -> 0.1799 | Helped |
| Add cluster association and tracking across intermediate scans | ICP-Flow · Waymo, dynamic foreground, up to 0.4 seconds | Acc-S (%): 67.59 -> 58.98 | Hurt |
Surprising findings
- Ours+FNN is less accurate than Ours on dynamic foreground yet is top among real-time methods, showing pseudo-label distillation keeps most performance.
- Learning-free ICP-Flow outperforms supervised baselines such as PCA and FastFlow on dynamic foreground in Waymo.
Limitations
- Feature-engineering design uses only geometry; fails with over/under-segmentation from ground removal and clustering, similar nearby objects, or sparse clusters.
- Naive cluster association does not enforce one-to-one correspondence, so a cluster may be matched multiple times.
- The rigid-body assumption fails for deformable or articulated objects such as buses and trucks.
Experimental setup
Datasets: Waymo nuScenes Argoverse-v2 KITTI
Benchmarks: Waymo Argoverse-v2 nuScenes
Metrics: EPE Acc-S Acc-R runtime
Baselines: NSFP Chodosh et al. FastNSF ZeroFlow RigidFlow FastFlow PCA
Builds on: Iterative Closest Point ZeroFlow Patchwork++ HDBSCAN KISS-ICP
Closest papers by meaning
- VoteFlow: Enforcing Local Rigidity in Self-Supervised Scene Flow (CVPR 2025 · similarity 0.90)
- Motion Cues from Image-based Point Tracking for LiDAR Scene Flow Estimation (NeurIPS 2026 · similarity 0.89)
- 3DSFLabelling: Boosting 3D Scene Flow Estimation by Pseudo Auto-labelling (CVPR 2024 · similarity 0.89)
- HiLoTs: High-Low Temporal Sensitive Representation Learning for Semi-Supervised LiDAR Segmentation in Autonomous Driving (CVPR 2025 · similarity 0.87)
- ZeroFlow: Scalable Scene Flow via Distillation (ICLR 2024 · similarity 0.87)
- Zero-Shot Monocular Scene Flow Estimation in the Wild (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.