Research map / Detection and segmentation / 3D detection and driving
Commonsense Prototype for Outdoor Unsupervised 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 CPD, an unsupervised 3D object detector that builds commonsense prototypes from high-confidence pseudo-labels to refine size/position of incomplete-object labels and guide sparse-object learning on LiDAR sequences.
Problem
Unsupervised 3D object detection from LiDAR suffers from pseudo-labels with wrong size and position for sparsely scanned, incomplete objects, which degrades self-training.
Approach
Construct a Commonsense Prototype (CProto) set of accurate boxes and dense points from complete objects using an unsupervised Completeness and Size Similarity score, then use it to regularize box sizes/positions of all pseudo-labels and to align a sparse-input detector with a dense-prototype network.
Key results
- On WOD validation, CPD reaches 57.79 vehicle 3D AP L1 at IoU0.5 vs 30.48 for OYSTER.
- Trained on WOD and tested on KITTI, CPD attains 90.85% and 81.01% car 3D AP@0.5 on easy and moderate, close to fully supervised CenterPoint (97.07/89.23).
- On PandaSet, CPD improves BEV AP at IoU0.7 from 18.1 (OYSTER) to 24.6.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add MFC to single-frame clustering (ablation) | CPD pseudo-label generation on WOD · WOD validation, Vehicle | 3D AP L1, IoU0.5: 17.36 -> 19.91 | Helped |
| add CBR to MFC (ablation) | CPD on WOD · WOD validation, Vehicle | 3D AP L2, IoU0.7: 4.77 -> 24.04 | Helped |
| add CST to MFC + CBR (ablation) | CPD on WOD · WOD validation, Vehicle | 3D AP L2, IoU0.7: 24.04 -> 32.13 | Helped |
| use 11 consecutive frames instead of fewer frames (ablation) | MFC pseudo-label generation · WOD validation | BEV recall and precision | Mixed |
| add MLO scoring to distance scoring (ablation) | CSS label scoring on WOD · WOD validation | BEV AP, IoU0.7: 14.80 -> 15.91 | Helped |
| add SS scoring to distance + MLO scoring (ablation) | CSS label scoring on WOD · WOD validation | BEV AP, IoU0.7: 15.91 -> 19.49 | Helped |
| add re-sizing to MFC labels (ablation) | CBR pseudo-label refinement on WOD · WOD validation | BEV recall, IoU0.5: 26.88 -> 30.79 | Helped |
| add re-localization to MFC + re-sizing (ablation) | CBR pseudo-label refinement on WOD · WOD validation | BEV recall, IoU0.5: 30.79 -> 43.47 | Helped |
| add CSS-weighted detection loss (ablation) | CST on WOD · WOD validation, Vehicle | 3D AP L2, IoU0.7: 24.04 -> 28.83 | Helped |
| add feature contrast loss (ablation) | CST on WOD · WOD validation, Vehicle | 3D AP L2, IoU0.7: 28.83 -> 29.58 | Helped |
| add box contrast loss (ablation) | CST on WOD · WOD validation, Vehicle | 3D AP L2, IoU0.7: 29.58 -> 32.13 | Helped |
| use CPD instead of OYSTER | unsupervised detection on PandaSet · PandaSet, 0–80m | BEV AP, IoU0.7: 18.1 -> 24.6 | Helped |
Surprising findings
- Cyclist detection remains very low (e.g., 21.91 L1 AP at IoU0.3 on WOD validation), a limitation attributed to scarce minority-class instances.
Limitations
- Much lower AP for minority classes such as cyclists due to scarce instances.
- Relies on predefined class-specific size thresholds and template boxes from commonsense knowledge.
- Evaluated mainly with single traversal and on specific datasets; generalization beyond them is not explored.
Experimental setup
Datasets: Waymo Open Dataset PandaSet KITTI
Benchmarks: Waymo Open Dataset PandaSet KITTI
Metrics: 3D AP BEV AP BEV recall 3D recall 3D precision
Baselines: DBSCAN MODEST OYSTER Proto-vanilla CenterPoint Sparsely-supervised 3D detector (2% labels)
Builds on: CenterPoint Voxel R-CNN OYSTER MODEST Prototypical VoteNet
Closest papers by meaning
- UNION: Unsupervised 3D Object Detection using Object Appearance-based Pseudo-Classes (NeurIPS 2024 · similarity 0.87)
- Unsupervised Object Detection With LIDAR Clues (CVPR 2021 · similarity 0.87)
- Training an Open-Vocabulary Monocular 3D Detection Model without 3D Data (NeurIPS 2024 · similarity 0.85)
- SAFDNet: A Simple and Effective Network for Fully Sparse 3D Object Detection (CVPR 2024 · similarity 0.85)
- STONE: A Submodular Optimization Framework for Active 3D Object Detection (NeurIPS 2024 · similarity 0.85)
- NBA3D: Neighbor-Based Confidence Adjustment for 3D Rare Object Detection Using LiDAR (AAAI 2025 · similarity 0.85)
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.