Research map / Detection and segmentation / 3D detection and driving
Uncertainty Meets Diversity: A Comprehensive Active Learning Framework for Indoor 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 the first active learning framework for indoor 3D object detection, combining an epistemic uncertainty score (localization-aware entropy plus a learned undetected-object counter) with a class-aware adaptive prototype bank for diversity, reaching ~85% of fully-supervised mAP with 10% of annotations.
Problem
Active learning for indoor 3D object detection is unexplored; indoor datasets have fewer samples per class, more classes, severe imbalance, and diverse scene types and intra-class variance, which make uncertainty and diversity estimation hard.
Approach
Combines a localization-aware uncertainty score (IoU between original and perturbed detections weighting Shannon entropy) with a learned estimate of undetected object counts, and selects diverse samples by jointly maximizing entropy and L1 norm of prototype histograms from a Class-aware Adaptive Prototype (CAP) bank built with an infinite-mixture-prototype style assignment.
Key results
- On SUN RGB-D with 10% annotation budget, the method reaches 56.11% mAP@0.25 (87.24% of fully-supervised), outperforming PPAL (52.55%).
- On ScanNetV2 with 10% budget, the method reaches 61.77% mAP@0.25 and 45.93% mAP@0.50, about 85% of fully-supervised performance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply UMD instead of PPAL | CAGroup3D active learning · SUN RGB-D (10% budget) | mAP@0.25: 52.55 -> 56.11 | Helped |
| apply UMD instead of PPAL | CAGroup3D active learning · SUN RGB-D (10% budget) | mAP@0.50: 33.51 -> 35.42 | Helped |
| apply UMD instead of PPAL | CAGroup3D active learning · ScanNetV2 (10% budget) | mAP@0.25: 59.44 -> 61.77 | Helped |
| apply UMD instead of PPAL | CAGroup3D active learning · ScanNetV2 (10% budget) | mAP@0.50: 43.15 -> 45.93 | Helped |
| add IoU weighting to diversity-only selection (ablation) | UMD · SUN RGB-D and ScanNetV2 | mAP@0.25: 54.5 -> 55.6; 58.7 -> 60.8 | Helped |
| add undetected-object uncertainty to diversity-only selection (ablation) | UMD · SUN RGB-D and ScanNetV2 | mAP@0.25: 54.5 -> 55.2; 58.7 -> 60.1 | Helped |
| replace CAP bank with fixed-number K-means prototypes (ablation) | UMD diversity selection · SUN RGB-D and ScanNetV2 | mAP@0.25: 56.1 -> 55.3; 61.8 -> 60.4 | Hurt |
| replace decomposed diversity optimization with greedy search on the original objective (ablation) | UMD diversity selection · SUN RGB-D and ScanNetV2 | mAP@0.25: 56.1 -> 54.5; 61.8 -> 58.7 | Hurt |
| replace Shannon entropy with probability margin (ablation) | UMD uncertainty estimation · SUN RGB-D and ScanNetV2 | mAP@0.25: 56.1 -> 55.8; 61.8 -> 61.5 | No clear effect |
| replace Shannon entropy with MC-Dropout uncertainty (ablation) | UMD uncertainty estimation · SUN RGB-D and ScanNetV2 | mAP@0.25: 56.1 -> 56.0; 61.8 -> 61.6 | No clear effect |
| vary uncertainty normalization scale k (ablation) | UMD on SUN RGB-D · SUN RGB-D | mAP@0.25 | Mixed |
| vary candidate-pool expansion ratio δ (ablation) | UMD on SUN RGB-D · SUN RGB-D | mAP@0.25 | Mixed |
Surprising findings
- Ablation shows the fixed-prototype K-means variant drops 2.4% mAP@0.50 on ScanNetV2 relative to the adaptive CAP bank.
- Greedy search on the original joint objective (Eq. 8) yields only minor gains, suggesting the decomposed problem restricts search to good subspaces.
Experimental setup
Datasets: SUN RGB-D ScanNetV2
Benchmarks: SUN RGB-D ScanNetV2
Metrics: mAP@0.25 mAP@0.50 mAR
Baselines: RAND MC-Dropout CORESET BADGE CDAL DivProto PPAL CRB KECOR
Builds on: CAGroup3D infinite mixture prototypes BADGE PPAL
Closest papers by meaning
- Cubify Anything: Scaling Indoor 3D Object Detection (CVPR 2025 · similarity 0.89)
- Plug and Play Active Learning for Object Detection (CVPR 2024 · similarity 0.88)
- UniDet3D: Multi-dataset Indoor 3D Object Detection (AAAI 2025 · similarity 0.88)
- MVSDet: Multi-View Indoor 3D Object Detection via Efficient Plane Sweeps (NeurIPS 2024 · similarity 0.86)
- SPGroup3D: Superpoint Grouping Network for Indoor 3D Object Detection (AAAI 2024 · similarity 0.86)
- ImOV3D: Learning Open Vocabulary Point Clouds 3D Object Detection from Only 2D Images (NeurIPS 2024 · similarity 0.86)
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.