Research map / Label-efficient and robust learning / Semi-supervised learning
Semi-supervised Class-Agnostic Motion Prediction with Pseudo Label Regeneration and BEVMix
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 a semi-supervised framework for class-agnostic motion prediction on LiDAR BEV sequences, combining mean-teacher self-training with a motion select-and-regenerate module (MSRM) for pseudo-label filtering and two augmentations (temporal sampling and BEVMix).
Problem
Class-agnostic motion prediction from LiDAR BEV sequences requires many manual motion annotations; existing annotation-efficient (self-/weakly-supervised) methods lag far behind fully supervised ones, and regression-based motion pseudo labels lack a confidence score for reliability filtering.
Approach
Adapts consistency-based self-training (mean-teacher/FixMatch-style weak-to-strong consistency) to motion prediction, judging pseudo-motion reliability via optimal-transport matching between warped current cells and future cells, re-generating unreliable labels from neighbours with a local-rigidity consistency check, and introducing sequence-aware augmentations (temporal sampling, BEVMix).
Key results
- With 1% labels, semi-supervised method reaches 1.9407 mean fast-speed error vs 3.5504 for self-supervised PillarMotion and 3.7242 for supervised-only at 1%.
- With 10% labels, semi-supervised mean fast error 1.2030 vs MotionNet fully supervised 1.0744 (small gap).
- BEVMix gives the largest gain in the ablation: fast mean error 2.9037 (mean-teacher) to 2.1427 with BEVMix alone, 1.9407 with all components.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use semi-supervised training with unlabeled data instead of supervised-only training | MotionNet · nuScenes, 1% labeled, fast | mean error: 3.7242 to 1.9407 | Helped |
| add temporal sampling to Mean-Teacher (ablation) | Mean-Teacher · nuScenes, 1% labeled, fast | mean error: 3.0868 to 2.9037 | Helped |
| add temporal sampling to Mean-Teacher (ablation) | Mean-Teacher · nuScenes, 1% labeled, slow | mean error: 0.4218 to 0.4594 | Hurt |
| add MSRM to Mean-Teacher + temporal sampling (ablation) | Mean-Teacher + temporal sampling · nuScenes, 1% labeled, slow | mean error: 0.4594 to 0.3946 | Helped |
| add MSRM to Mean-Teacher + temporal sampling (ablation) | Mean-Teacher + temporal sampling · nuScenes, 1% labeled, fast | mean error: 2.9037 to 2.7705 | Helped |
| add BEVMix to Mean-Teacher + temporal sampling + MSRM (ablation) | Mean-Teacher + temporal sampling + MSRM · nuScenes, 1% labeled, fast | mean error: 2.7705 to 1.9407 | Helped |
| add pseudo-label selection (ablation) | Mean-Teacher · nuScenes, 1% labeled, fast | mean error: 2.9037 to 2.8689 | Helped |
| regenerate pseudo labels from neighbors without local consistency evaluation (ablation) | Mean-Teacher with pseudo-label selection · nuScenes, 1% labeled, fast | mean error: 2.8689 to 4.0773 | Hurt |
| add consistency evaluation when re-generating pseudo labels (ablation) | Mean-Teacher with pseudo-label selection and regeneration · nuScenes, 1% labeled, fast | mean error: 4.0773 to 2.7705 | Helped |
| use Mixup (ablation) | Mean-Teacher · nuScenes, 1% labeled, static | mean error: 0.0162 to 0.0106 | Helped |
| use Mixup (ablation) | Mean-Teacher · nuScenes, 1% labeled, fast | mean error: 2.9037 to 2.8902 | No clear effect |
| use BEVMix instead of CutMix (ablation) | Mean-Teacher · nuScenes, 1% labeled, fast | mean error: 2.6157 to 2.1427 | Helped |
Surprising findings
- Mixup reduces static error strongly (0.0162 to 0.0106) but barely improves fast-speed error, while CutMix helps fast error more than Mixup.
Limitations
- Gap to fully supervised methods remains at fast speeds and at the 1% setting in some metrics
- Local rigidity assumption for re-generating pseudo labels applies mainly to rigid objects such as cars
- Temporal sampling makes samples harder to learn and slightly hurts slow-speed error
Experimental setup
Datasets: nuScenes
Benchmarks: nuScenes
Metrics: mean L2 displacement error median L2 displacement error
Baselines: LSTM-ED MotionNet BE-STI PillarMotion WeakMotionNet Mean-Teacher Mixup CutMix
Builds on: MotionNet FixMatch Mean Teacher Mixup CutMix
Closest papers by meaning
- Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations (CVPR 2024 · similarity 0.89)
- SmartPretrain: Model-Agnostic and Dataset-Agnostic Representation Learning for Motion Prediction (ICLR 2025 · similarity 0.88)
- Uncertainty-Guided Never-Ending Learning to Drive (CVPR 2024 · similarity 0.86)
- Revisiting Semi-Supervised Learning in the Era of Foundation Models (NeurIPS 2025 · similarity 0.84)
- Semi-Supervised Vision-Centric 3D Occupancy World Model for Autonomous Driving (ICLR 2025 · similarity 0.83)
- Towards Realistic Model Selection for Semi-supervised Learning (ICML 2024 · similarity 0.83)
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.