Research map / Detection and segmentation / Weakly supervised detection and segmentation
A-Teacher: Asymmetric Network for 3D Semi-Supervised Object Detection
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-Teacher, an online asymmetric semi-supervised framework for LiDAR 3D object detection that attaches an attention-based temporal refinement model to a vanilla teacher so pseudo labels benefit from past and future frames while the teacher is updated by EMA.
Problem
Semi-supervised LiDAR 3D object detection: symmetric online teacher-student methods have weak teachers, while offline asymmetric methods with complex refinement (e.g., trackers) cannot be jointly trained with the student and teacher.
Approach
A lightweight attention-based refinement model (propagation-based box aggregation, dreaming-based box aggregation, spatio-temporal deformable aggregation) that uses supporting frames without a tracker, so the asymmetric teacher can be trained jointly and updated online via EMA.
Key results
- On Waymo 1% labeled setting, A-Teacher reaches 46.6 L1 mAP versus 41.9 for HSSDA and 27.7 for the PV-RCNN supervised baseline.
- Uses 71.1 V100 GPU hours versus 84.6 for HSSDA (about 15.9% less training resources).
- On ONCE (small setting), mAP improves from 51.89 to 60.52 (+8.63) over the labeled-only baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the A-Teacher refinement framework to PV-RCNN | PV-RCNN · Waymo, 1% labeled, L1 | mAP: 27.7 -> 46.6 | Helped |
| Add the A-Teacher framework with 4% unlabeled data | Second · Waymo, Vehicle L1 | mAP: 39.8 -> 43.6 | Helped |
| Add the A-Teacher framework with 4% unlabeled data | Voxel-RCNN · Waymo, Vehicle L1 | mAP: 50.6 -> 55.4 | Helped |
| Add the A-Teacher framework to PV-RCNN 4-frame | PV-RCNN 4-frame · Waymo, Vehicle L1 | mAP: 49.8 -> 55.8 | Helped |
| Use A-Teacher instead of the labeled-only baseline | ONCE small setting · ONCE | mAP: 51.89 -> 60.52 | Helped |
| Remove the refinement model (ablation) | Online symmetric A-Teacher · Waymo, 1% labeled + 4% unlabeled, Vehicle L1 | mAP: 54.4 -> 52.4 | Hurt |
| Disable teacher EMA updates (ablation) | A-Teacher · Waymo, 1% labeled + 4% unlabeled, Vehicle L1 | mAP: 54.4 -> 52.9 | Hurt |
| Add PBA (ablation) | A-Teacher refinement model · Waymo, 1% labeled + 4% unlabeled, Vehicle L1 | mAP: 53.0 -> 54.4 | Helped |
| Add DBA (ablation) | A-Teacher refinement model · Waymo, 1% labeled + 4% unlabeled, Vehicle L1 | mAP: 54.1 -> 54.4 | Helped |
| Add STA (ablation) | A-Teacher refinement model · Waymo, 1% labeled + 4% unlabeled, Vehicle L1 | mAP: 53.5 -> 54.4 | Helped |
| Increase unlabeled data from 4% to 19% | A-Teacher with PV-RCNN · Waymo, L1 | mAP: 37.4 -> 43.0 | Helped |
| Increase unlabeled data from 19% to 99% | A-Teacher with PV-RCNN · Waymo, L1 | mAP: 43.0 -> 46.6 | Helped |
Surprising findings
- Gains for pedestrian and cyclist categories are much larger than for vehicles (e.g., cyclist L1 mAP from 4.3 to 33.7), which the authors attribute possibly to class imbalance.
Limitations
- Only k=1 supporting frames (previous and next timestamp) used for training efficiency
- Gains on pedestrian and cyclist are attributed to possible class imbalance and left for future study
- Compared mostly against PV-RCNN-based methods; comparison to HSSDA involves differing TTA settings
Experimental setup
Datasets: Waymo Open Dataset ONCE
Benchmarks: Waymo Open Dataset ONCE
Metrics: AP APH mAP mAPH
Baselines: PV-RCNN DetMatch HSSDA NoiseDet Second Voxel-RCNN
Builds on: PV-RCNN HSSDA Mean Teacher (EMA) Voxel RoI pooling Voxel-RCNN
Closest papers by meaning
- Semi-supervised 3D Object Detection with PatchTeacher and PillarMix (AAAI 2024 · similarity 0.92)
- Beyond Virtual Points: Depth-Enhanced LiDAR-only 3D Object Detection with Semi-Supervised Learning (Student Abstract) (AAAI 2025 · similarity 0.88)
- Weak-to-Strong 3D Object Detection with X-Ray Distillation (CVPR 2024 · similarity 0.88)
- Leveraging Imagery Data with Spatial Point Prior for Weakly Semi-supervised 3D Object Detection (AAAI 2024 · similarity 0.88)
- Dual-Perspective Knowledge Enrichment for Semi-supervised 3D Object Detection (AAAI 2024 · similarity 0.88)
- STONE: A Submodular Optimization Framework for Active 3D Object Detection (NeurIPS 2024 · 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.