Research map / Point clouds / Point cloud classification

3DSFLabelling: Boosting 3D Scene Flow Estimation by Pseudo Auto-labelling

CVPR 2024 · Point clouds › Point cloud classification · approach: LiDAR point cloud compression and completion · 9 citations (OpenAlex)

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.

Generates pseudo 3D scene flow labels for unlabelled real LiDAR point clouds by optimizing rigid box motion parameters and warping source clouds with global-local motion augmentation, then training existing scene flow networks on them without any ground-truth labels.

Problem

Learning 3D scene flow from LiDAR point clouds suffers from poor generalization from synthetic datasets to real driving scenes, scarce real-world scene flow labels, and poor performance on sparse real LiDAR data.

Approach

Treat scene flow as rigid object-level box motion: optimize anchor boxes with global ego-motion and per-box motion parameters and motion probabilities against nearest-neighbour losses on unlabelled point clouds, then warp source points to synthesize target clouds whose correspondences give pseudo flow labels; augment global and local motion (Lie-algebra rotation noise) to diversify labels.

Key results

• On LiDAR KITTI, GMSF+3DSFlabelling reduces EPE3D from 0.1900m to 0.0078m (Acc3DS 0.9924), outperforming all prior supervised and unsupervised methods without ground-truth labels.
• On Argoverse, GMSF+3DSFlabelling achieves EPE3D 0.0093m vs 7.2776m for the baseline GMSF; on nuScenes, 0.0185m vs 9.4231m.
• FLOT+3DSFlabelling with 0.1M parameters and 78 ms runtime surpasses all listed supervised, self-supervised, weakly supervised and optimization-based methods on Argoverse (EPE3D 0.0107m).

Measured effects reported in the paper

Surprising findings

• Models trained on the auto-generated labels alone outperform all prior supervised models trained on FT3D or real labels, despite using no ground-truth scene flow labels.
• Standard random-noise augmentation is less helpful than the proposed motion-level augmentation, which applies motion noise only to dynamic boxes.

Limitations

• Evaluation LiDAR KITTI ground truth contains errors, so the authors fuse ground truth with the first frame to create a corrected second frame.
• Relies on coarse normals from Open3D and initial anchor boxes that must be set according to point range; hyperparameters such as motion-probability slope and threshold are manually set.

Experimental setup

Datasets: KITTI Odometry nuScenes sweeps Argoverse 2 Sensor Dataset LiDAR KITTI scene flow Argoverse scene flow nuScenes scene flow

Benchmarks: LiDAR KITTI scene flow Argoverse scene flow nuScenes scene flow

Metrics: EPE3D Acc3DS Acc3DR Outliers Params inference time

Baselines: Graph prior RSF NSFP R3DSF FlowNet3D PointPWC DCA-SRSFE FLOT MSBRN GMSF Mittal et al. Jiang et al.

Builds on: RSF RigidFlow GMSF FLOT MSBRN

Closest papers by meaning

• Motion Cues from Image-based Point Tracking for LiDAR Scene Flow Estimation (NeurIPS 2026 · similarity 0.89)
• ICP-Flow: LiDAR Scene Flow Estimation with ICP (CVPR 2024 · similarity 0.89)
• ZeroFlow: Scalable Scene Flow via Distillation (ICLR 2024 · similarity 0.88)
• VoteFlow: Enforcing Local Rigidity in Self-Supervised Scene Flow (CVPR 2025 · similarity 0.87)
• SeMoLi: What Moves Together Belongs Together (CVPR 2024 · similarity 0.87)
• Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations (CVPR 2024 · similarity 0.86)