Research map / Point clouds / Point cloud classification
Instantaneous Perception of Moving Objects in 3D
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 S'More, a LiDAR-based framework that uses local occupancy completion of object point clouds to detect and estimate subtle vehicle motions instantaneously, and contributes a small-motion benchmark from Waymo.
Problem
Detecting and quantifying subtle, instantaneous motions of vehicles from sparse LiDAR point clouds, where the swimming effect (apparent motion of static objects due to lack of cross-frame point correspondences) confounds true small motion.
Approach
Learn local occupancy completion of the visible object surface (supervised by points from neighboring frames warped with GT motion, plus line-of-sight free space) end-to-end with a static/moving classifier and a flow predictor, densifying shape cues to mitigate swimming artifacts; trained only on small-motion samples.
Key results
- S'More achieves EPE 0.0437, angle error 0.3189 and F1 0.8323 on the small-motion test set, beating the best baseline ICP (F1 0.7456) and CenterPoint (F1 0.7270).
- Removing occupancy completion lowers average F1 across five thresholds from 0.8100 to 0.7932 (w/o Occ loss), and w/o angular loss drops average angle error to 0.4066 vs 0.3758.
- On large motions, CenterPoint reaches F1 0.9998 vs S'More 0.9847, with CenterPoint giving lower EPE (0.1155 vs 0.2790).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use S’More instead of the strongest reported baseline | Subtle-motion perception · Subtle-motion benchmark | EPE: 0.0554 -> 0.0437 | Helped |
| Use S’More instead of the strongest reported baseline | Subtle-motion perception · Subtle-motion benchmark | Angle Error: 0.4170 -> 0.3189 | Helped |
| Use S’More instead of the strongest reported baseline | Subtle-motion perception · Subtle-motion benchmark | F1 Score: 0.7856 -> 0.8323 | Helped |
| Use S’More instead of CenterPoint | Large-motion perception · Large motions | EPE: 0.1155 -> 0.2790 | Hurt |
| Use S’More instead of CenterPoint | Large-motion perception · Large motions | F1 Score: 0.9998 -> 0.9847 | Hurt |
| Include large-motion samples in training (ablation) | S’More · Subtle-motion benchmark | EPE: 0.0437 -> 0.0979 | Hurt |
| Include large-motion samples in training (ablation) | S’More · Subtle-motion benchmark | Angle Error: 0.3189 -> 0.5550 | Hurt |
| Include large-motion samples in training (ablation) | S’More · Subtle-motion benchmark | F1 Score: 0.8323 -> 0.1414 | Hurt |
| Add occupancy completion loss Locc (ablation) | S’More · Subtle-motion benchmark, average over five thresholds | EPE: 0.0461 -> 0.0429 | Helped |
| Add occupancy completion loss Locc (ablation) | S’More · Subtle-motion benchmark, average over five thresholds | Angle Error: 0.3879 -> 0.3758 | Helped |
| Add occupancy completion loss Locc (ablation) | S’More · Subtle-motion benchmark, average over five thresholds | F1 Score: 0.7932 -> 0.8100 | Helped |
| Add angular loss Langle (ablation) | S’More · Subtle-motion benchmark, average over five thresholds | Angle Error: 0.4066 -> 0.3758 | Helped |
Surprising findings
- Training with large-motion data included severely degrades small-motion performance (F1 0.1414 vs 0.8323 with small-motion-only training).
- False positives (static objects detected as moving, from swimming effect) occur far more often than false negatives.
Limitations
- Struggles with extremely sparse or heavily occluded objects.
- Only handles vehicles, not pedestrians or cyclists.
- Underperforms CenterPoint on large motions and in flow precision.
Experimental setup
Datasets: Waymo Open Dataset
Benchmarks: Waymo small-motion benchmark
Metrics: F1 score end-point error (EPE) angle error
Baselines: ICP Point-to-Plane ICP Generalized ICP FastNSF CenterPoint
Builds on: CenterPoint FastNSF Generalized ICP 4D occupancy forecasting (Khurana et al.)
Closest papers by meaning
- Beyond 3D Siamese Tracking: A Motion-Centric Paradigm for 3D Single Object Tracking in Point Clouds (CVPR 2022 · similarity 0.86)
- Towards Flexible 3D Perception: Object-Centric Occupancy Completion Augments 3D Object Detection (NeurIPS 2024 · similarity 0.86)
- CorrBEV: Multi-View 3D Object Detection by Correlation Learning with Multi-modal Prototypes (CVPR 2025 · similarity 0.86)
- PTT: Point-Trajectory Transformer for Efficient Temporal 3D Object Detection (CVPR 2024 · similarity 0.85)
- UAWTrack: Universal 3D Single Object Tracking in Adverse Weather (AAAI 2025 · similarity 0.85)
- UnO: Unsupervised Occupancy Fields for Perception and Forecasting (CVPR 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.