Research map / Detection and segmentation / 3D detection and driving
Pseudo Label Refinery for Unsupervised Domain Adaptation on Cross-dataset 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 PERE, a pseudo label refinery for cross-dataset 3D unsupervised domain adaptation that replaces or removes unreliable pseudo boxes, adds interpolated/extrapolated proposals, and aligns cross-domain RoI features via triplet loss.
Problem
Self-training for unsupervised domain adaptation of LiDAR 3D object detection selects pseudo boxes with a confidence threshold, leaving unreliable boxes whose points are ambiguous foreground/background, and cross-dataset point-number inconsistency (IPNI) between high-beam and low-beam datasets degra
Approach
Complementary augmentation probabilistically either removes an unreliable box's points or replaces the box with a high-confidence box from a database (weighted by its IoU confidence), combined with interpolation/extrapolation of proposals to counter point-number inconsistency and a cross-domain triplet loss to align RoI features.
Key results
- On N→K with PV-RCNN, PERE reaches 44.85 average AP3D versus 40.28 for ST3D++ and 65.72 oracle.
- On W→K with PV-RCNN, PERE reaches 52.63 average AP3D versus 49.41 for ST3D++.
- Ablation on N→K shows complementary augmentation gives the largest single gain (about 9.36% AP3D improvement over the naive self-training baseline).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the source-domain intra-domain triplet loss (ablation) | Self-training with PVRCNN · N → K (Car) | AP3D: 57.79 -> 58.21 | Helped |
| Add cross-domain RoI feature alignment (ablation) | Self-training with PVRCNN · N → K (Car) | AP3D: 57.79 -> 59.92 | Helped |
| Add interpolation and extrapolation proposal generation to ST + RFA (ablation) | Self-training with PVRCNN · N → K (Car) | AP3D: 59.92 -> 62.93 | Helped |
| Add complementary augmentation (ablation) | Self-training with PVRCNN · N → K (Car) | AP3D: 62.93 -> 68.34 | Helped |
| Add BoxReplace rather than directly removing unreliable boxes (ablation) | Complementary augmentation with PVRCNN · N → K (Car) | AP3D: 64.01 -> 68.34 | Helped |
| Add PointRemove (ablation) | Complementary augmentation with PVRCNN · N → K (Car) | AP3D: 64.32 -> 68.34 | Helped |
| Use weighted sampling instead of random sampling (ablation) | Complementary augmentation with PVRCNN · N → K (Car) | AP3D: 62.37 -> 68.34 | Helped |
| Use PERE rather than the strongest listed baseline | PVRCNN · N → K | Average AP3D: 40.28 -> 44.85 | Helped |
| Use PERE rather than the strongest listed baseline | PVRCNN · W → K | Average AP3D: 49.41 -> 52.63 | Helped |
| Use PERE rather than the strongest listed baseline | PVRCNN · W → N | Average AP3D: 13.15 -> 14.36 | Helped |
| Use PERE rather than the strongest listed baseline | SECOND · N → K | Average AP3D: 35.95 -> 40.57 | Helped |
| Use PERE rather than the strongest listed baseline | SECOND · W → K | Average AP3D: 41.58 -> 43.30 | Helped |
Limitations
- Multi-category setting is harder, so reported gains are lower than single-category results in prior work
- Dense proposal generation is computationally costly, motivating the limited two-proposal design
- Some AP3D results for pedestrians and cyclists are second-best rather than best
Experimental setup
Datasets: Waymo nuScenes KITTI
Benchmarks: KITTI nuScenes Waymo
Metrics: AP_BEV AP_3D Closed Gap
Baselines: Source Only ST3D ST3D++ DTS LD
Builds on: ST3D PV-RCNN SECOND-IoU OpenPCDet triplet loss
Closest papers by meaning
- ST3D: Self-Training for Unsupervised Domain Adaptation on 3D Object Detection (CVPR 2021 · similarity 0.91)
- Multi-Modal Assistance for Unsupervised Domain Adaptation on Point Cloud 3D Object Detection (AAAI 2026 · similarity 0.91)
- Unified Domain Generalization and Adaptation for Multi-View 3D Object Detection (NeurIPS 2024 · similarity 0.88)
- Dual-Perspective Knowledge Enrichment for Semi-supervised 3D Object Detection (AAAI 2024 · similarity 0.87)
- Active Domain Adaptation with False Negative Prediction for Object Detection (CVPR 2024 · similarity 0.86)
- UNION: Unsupervised 3D Object Detection using Object Appearance-based Pseudo-Classes (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.