Research map / Detection and segmentation / 3D detection and driving
Semi-supervised 3D Object Detection with PatchTeacher and PillarMix
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 PatchTeacher (high-resolution partial-scene teacher with Patch Normalizer, Quadrant Align, Fovea Selection) and PillarMix augmentation for semi-supervised 3D object detection on Waymo and ONCE.
Problem
Semi-supervised 3D object detection relies on pseudo-label quality, but complete-scene teachers cannot use high-resolution voxelization due to memory limits, and strong augmentation for pseudo-labeled LiDAR data is underexplored.
Approach
Train a teacher on sequentially processed patches of a scene so it can use a very small voxel size (3.5 cm), with normalization, quadrant alignment and overlapped-fovea inference to handle patch-specific issues; plus PillarMix, which cross-mixes pillars from different LiDAR scans so adjacent pillars come from different frames.
Key results
- On Waymo 5% labels, PTPM reaches 54.53 overall 3D AP/APH (L2) vs ProficientTeacher 51.10 mAP.
- On ONCE, PTPM achieves 60.75/61.47/62.68 mAP on small/medium/large unlabeled splits, beating ProficientTeacher by 3.03/1.58/1.28.
- PatchTeacher reaches about 10 mAPH improvement over the student on the Waymo unlabeled split (56.53/53.49 vs 46.20/41.40).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PTPM instead of ProficientTeacher | SECOND · Waymo 5% labeled | Overall 3D AP/APH @0.7 Level 2: 51.10/45.75 -> 54.53/51.07 | Helped |
| Use PTPM instead of ProficientTeacher | SECOND · Waymo 20% labeled | Overall 3D AP/APH @0.7 Level 2: 58.59/54.16 -> 60.49/57.02 | Helped |
| Use PTPM instead of ProficientTeacher | SECOND · Waymo 100% labeled | Overall 3D AP/APH @0.7 Level 2: 62.96/59.14 -> 65.73/62.13 | Helped |
| Use PTPM instead of ProficientTeacher | SECOND · ONCE validation, Small (100K unlabeled samples) | mAP: 57.72 -> 60.75 | Helped |
| Use PTPM instead of ProficientTeacher | SECOND · ONCE validation, Medium (500K unlabeled samples) | mAP: 59.89 -> 61.47 | Helped |
| Use PTPM instead of ProficientTeacher | SECOND · ONCE validation, Large (1M unlabeled samples) | mAP: 61.40 -> 62.68 | Helped |
| Add pseudo-labeling to the labeled-only baseline (ablation) | SECOND semi-supervised detector · Waymo 5% labeled | Overall 3D AP/APH @0.7 Level 2: 44.87/40.18 -> 48.63/43.53 | Helped |
| Add Semi-Sampling to pseudo-labeling (ablation) | SECOND semi-supervised detector · Waymo 5% labeled | Overall 3D AP/APH @0.7 Level 2: 48.63/43.53 -> 51.61/47.43 | Helped |
| Add PillarMix to pseudo-labeling and Semi-Sampling (ablation) | SECOND semi-supervised detector · Waymo 5% labeled | Overall 3D AP/APH @0.7 Level 2: 51.61/47.43 -> 52.88/48.75 | Helped |
| Add PatchTeacher to pseudo-labeling, Semi-Sampling, and PillarMix (ablation) | SECOND semi-supervised detector · Waymo 5% labeled | Overall 3D AP/APH @0.7 Level 2: 52.88/48.75 -> 55.81/52.17 | Helped |
| Add Patch Normalizer (ablation) | PatchTeacher · Waymo 5% labeled | Vehicle 3D AP/APH @0.7 Level 2: 51.83 / 51.26 -> 52.61 / 52.09 | Helped |
| Add Quadrant Align to Patch Normalizer (ablation) | PatchTeacher · Waymo 5% labeled | Vehicle 3D AP/APH @0.7 Level 2: 52.61 / 52.09 -> 55.41 / 54.89 | Helped |
Surprising findings
- Averaging the Patch Normalizer negative loss by the number of negatives drops performance drastically (mAP 48.64 vs 57.88).
- Using two PatchTeachers (close and far patches separately) does not improve over a shared teacher.
Limitations
- Patch-based teacher training is about 8x slower per epoch without filtering, requiring patch selection heuristics.
- Gains rely on many hyperparameters (alpha, delta, pillar size, score thresholds) tuned on Waymo 5%.
- Quadrant Align assumes radially distributed point clouds (rotational symmetry), which may not hold for all sensors.
Experimental setup
Datasets: Waymo Open Dataset ONCE
Benchmarks: Waymo Open Dataset ONCE
Metrics: 3D AP APH mAP mAPH
Baselines: FixMatch PseudoAugment ProficientTeacher DetMatch HSSDA 3DIoUMatch MeanTeacher LaserMix PolarMix ShuffleDA
Builds on: ProficientTeacher PseudoAugment SECOND LaserMix PolarMix
Closest papers by meaning
- A-Teacher: Asymmetric Network for 3D Semi-Supervised Object Detection (CVPR 2024 · similarity 0.92)
- Dual-Perspective Knowledge Enrichment for Semi-supervised 3D Object Detection (AAAI 2024 · similarity 0.89)
- Leveraging Imagery Data with Spatial Point Prior for Weakly Semi-supervised 3D Object Detection (AAAI 2024 · similarity 0.89)
- Beyond Virtual Points: Depth-Enhanced LiDAR-only 3D Object Detection with Semi-Supervised Learning (Student Abstract) (AAAI 2025 · similarity 0.88)
- Voxel or Pillar: Exploring Efficient Point Cloud Representation for 3D Object Detection (AAAI 2024 · similarity 0.88)
- Leveraging Temporal Cues for Semi-Supervised Multi-View 3D Object Detection (CVPR 2025 · similarity 0.88)
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.