Research map / Point clouds / Segmentation and detection
STONE: A Submodular Optimization Framework for Active 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 STONE, a two-stage submodular optimization framework for active 3D object detection that selects point clouds for representativeness (gradient-based submodular subset selection with class-reweighted losses) and label-distribution balance (entropy-based class balancing).
Problem
Active learning for LiDAR-based 3D object detection is costly to label; existing methods ignore class imbalance across labeled point clouds and fail to cover varying difficulty levels.
Approach
Formulate active 3D detection sample selection as submodular optimization with two objectives: a feature-based submodular coverage function over MC-dropout gradients (with class-balanced reweighted detection losses) and an entropy-based submodular function that balances the cumulative label distribution of the labeled set.
Key results
- On KITTI with PV-RCNN at 1% queried boxes, STONE reaches average 3D AP of 82.17 (easy), 68.75 (moderate), 64.05 (hard), above KECOR (81.63, 68.67, 63.42).
- With SECOND backbone on KITTI, STONE improves HARD 3D mAP to 58.75 versus 55.34 for KECOR.
- On VOC07 in the first query (2k images), STONE reaches 65.34% mAP versus 62.91% for the best AL-MDN variant.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use STONE instead of KECOR | PV-RCNN active 3D detection · KITTI validation, HARD, 1% queried boxes | 3D AP (%): 63.42 -> 64.05 | Helped |
| Use STONE instead of KECOR | SECOND active 3D detection · KITTI validation, HARD, 1% queried boxes | 3D mAP (%): 55.34 -> 58.75 | Helped |
| Use STONE instead of AL-MDN_eff | SSD with VGG-16 · VOC07 test, 1st query | mAP (%): 62.91±0.16 -> 65.34±0.34 | Helped |
| Use STONE instead of AL-MDN_eff | SSD with VGG-16 · VOC07 test, 2nd query | mAP (%): 67.61±0.17 -> 67.01±0.47 | Hurt |
| Use STONE instead of AL-MDN_eff | SSD with VGG-16 · VOC07 test, 3rd query | mAP (%): 69.66±0.17 -> 69.03±0.55 | Hurt |
| Add regression and classification loss reweighing (ablation) | STONE active 3D detection · KITTI validation, EASY | 3D AP (%): 80.81 -> 83.03 | Helped |
| Add regression and classification loss reweighing (ablation) | STONE active 3D detection · KITTI validation, MODERATE | 3D AP (%): 68.78 -> 70.33 | Helped |
| Add regression and classification loss reweighing (ablation) | STONE active 3D detection · KITTI validation, HARD | 3D AP (%): 62.88 -> 65.33 | Helped |
| Add regression-loss reweighing (ablation) | STONE active 3D detection · KITTI validation, HARD | 3D AP (%): 63.59 -> 65.33 | Helped |
| Add classification-loss reweighing (ablation) | STONE active 3D detection · KITTI validation, HARD | 3D AP (%): 62.93 -> 65.33 | Helped |
| Add SDMCB to GBSSS (ablation) | STONE active 3D detection · KITTI validation, EASY | 3D AP (%): 79.60 -> 83.03 | Helped |
| Add SDMCB to GBSSS (ablation) | STONE active 3D detection · KITTI validation, MODERATE | 3D AP (%): 66.50 -> 70.33 | Helped |
Surprising findings
- KECOR marginally surpasses STONE at 2% and 3% labeled bounding boxes on KITTI HARD, which the authors attribute to STONE selecting scenes with more objects and hitting the box budget early.
- On VOC07, STONE does not beat AL-MDN in the 2nd and 3rd queries (67.01 and 69.03 vs 67.61 and 69.66).
Limitations
- Does not reduce the running time of existing active 3D detection methods, due to the large pool of unlabeled point clouds.
- At 2% and 3% labeled boxes on KITTI HARD, KECOR marginally outperforms STONE because STONE selects scenes with many objects and exhausts the box budget early.
Experimental setup
Datasets: KITTI Waymo Open Dataset PASCAL VOC
Benchmarks: KITTI Waymo Open Dataset PASCAL VOC07
Metrics: 3D AP BEV AP APH mAP
Baselines: RANDOM ENTROPY LLAL CORESET BADGE MC-MI MC-REG LT/C CONSENSUS CRB KECOR AL-MDN
Builds on: PV-RCNN CRB KECOR MC-dropout BADGE
Closest papers by meaning
- Exploring Active 3D Object Detection from a Generalization Perspective (ICLR 2023 · similarity 0.93)
- AS-Det: Active Sampling for Adaptive 3D Object Detection in Point Clouds (AAAI 2025 · similarity 0.88)
- iDet3D: Towards Efficient Interactive Object Detection for LiDAR Point Clouds (AAAI 2024 · similarity 0.87)
- SAFDNet: A Simple and Effective Network for Fully Sparse 3D Object Detection (CVPR 2024 · similarity 0.87)
- Beyond Virtual Points: Depth-Enhanced LiDAR-only 3D Object Detection with Semi-Supervised Learning (Student Abstract) (AAAI 2025 · similarity 0.87)
- A-Teacher: Asymmetric Network for 3D Semi-Supervised Object Detection (CVPR 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.