Research map / Detection and segmentation / Weakly supervised detection and segmentation
Dual-Perspective Knowledge Enrichment for Semi-supervised 3D 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 DPKE, a semi-supervised 3D indoor object detector that enriches data with class-probability-weighted instance insertion and matches student-teacher proposal features weighted by Chamfer-distance geometry similarity.
Problem
Semi-supervised 3D object detection in cluttered indoor scenes suffers from limited unlabeled data diversity and a trade-off between pseudo-label quality and recall (strict thresholds lower recall, loose consistency gives noisy labels).
Approach
Enrich knowledge from two perspectives: a class-probabilistic data augmentation that pastes instances from a proposal bank with class-dependent sampling probabilities derived from logits, and a geometry-aware feature matching loss that aligns student-teacher proposal features on low-confidence proposals, weighted by Chamfer-distance-based geometric similarity.
Key results
- On ScanNet 10% labels, DPKE reaches 51.88 mAP@0.25 and 34.06 mAP@0.5, versus 47.53 / 28.63 for 3DIoUMatch (reproduced) and 48.32 / 30.15 for Semi-Sampling.
- On SUN RGB-D 5% labels, DPKE reaches 41.51 mAP@0.25 and 24.98 mAP@0.5, improving over Semi-Sampling (39.79 / 22.60).
- Gains over the previous state of the art are about 2-4 points on ScanNet across label ratios and about 2 points on SUN RGB-D; at 100% labels the gain is only about 1 point.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DPKE rather than Semi-Sampling | IoU-aware VoteNet · ScanNet 10% | mAP@0.25: 48.32 -> 51.88 | Helped |
| Use class-probability sampling rather than uniform sampling (ablation) | DPKE · ScanNet 10% | mAP@0.25: 48.32 -> 50.63 | Helped |
| Use low-logit sampling rather than uniform sampling (ablation) | class-probabilistic augmentation · ScanNet 10% | mAP@0.25: 48.32 -> 48.07 | Hurt |
| Use low-logit sampling rather than uniform sampling (ablation) | class-probabilistic augmentation · SUN RGB-D 5% | mAP@0.25: 39.79 -> 39.42 | Hurt |
| Add feature matching without geometry weights (ablation) | DPKE · ScanNet 10% | mAP@0.25: 47.53 -> 48.59 | Helped |
| Add geometry weights to feature matching (ablation) | DPKE · ScanNet 10% | mAP@0.25: 48.59 -> 49.68 | Helped |
| Use geometry-aware weights instead of constant weights (ablation) | feature matching · SUN RGB-D 5% | mAP@0.25: 39.19 -> 39.44 | Helped |
| Add SESS consistency loss to 3DIoUMatch (ablation) | 3DIoUMatch · ScanNet 5% | mAP@0.25: 39.44 -> 38.85 | Hurt |
| Add SESS consistency loss to 3DIoUMatch (ablation) | 3DIoUMatch · SUN RGB-D 5% | mAP@0.25: 38.72 -> 37.04 | Hurt |
| Use HLS rather than uniform sampling (ablation) | class-probabilistic augmentation · ScanNet 10% | mAP@0.25: 48.32 -> 50.63 | Helped |
| Use geometry-aware weights rather than HCD weights (ablation) | feature matching · ScanNet 10% | mAP@0.25: 48.83 -> 49.68 | Helped |
Surprising findings
- Sampling instances from classes with lower logits (harder classes) harmed mAP, so the class-probability weighting favors well-learned classes.
- Naively adding SESS's consistency loss to 3DIoUMatch lowered performance on both datasets.
Limitations
- Gains on SUN RGB-D are smaller, attributed to lower-quality point clouds and reliance on ground-truth for the proposal bank.
- Geometry-aware feature matching is inactive in fully labeled settings, giving only about 1% gain at 100% labels.
- Unlabeled-scene collision detection relies on predictions from the pretrained model, and the geometry weight is costly without subsampling.
Experimental setup
Datasets: ScanNet SUN RGB-D
Benchmarks: ScanNet SUN RGB-D
Metrics: mAP@0.25 mAP@0.5
Baselines: VoteNet SESS 3DIoUMatch Confid-3DIoUMatch Semi-Sampling
Builds on: Mean Teacher 3DIoUMatch SESS semi-sampling VoteNet
Closest papers by meaning
- Semi-supervised 3D Object Detection with PatchTeacher and PillarMix (AAAI 2024 · similarity 0.89)
- Leveraging Temporal Cues for Semi-Supervised Multi-View 3D Object Detection (CVPR 2025 · similarity 0.89)
- Decoupled Pseudo-labeling for Semi-Supervised Monocular 3D Object Detection (CVPR 2024 · similarity 0.88)
- A-Teacher: Asymmetric Network for 3D Semi-Supervised Object Detection (CVPR 2024 · similarity 0.88)
- SP3D: Boosting Sparsely-Supervised 3D Object Detection via Accurate Cross-Modal Semantic Prompts (CVPR 2025 · similarity 0.87)
- Unified Domain Generalization and Adaptation for Multi-View 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.