Research map / Detection and segmentation / 3D detection and driving
Towards Learning Group-Equivariant Features for Domain Adaptive 3D 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 GroupExp-DA, a self-training domain adaptation add-on for LiDAR 3D detection that learns GMM-based object groups and uses group-equivariant spatial features plus an explorative group update to reduce bias and false negatives.
Problem
Domain adaptation for 3D object detection in outdoor point clouds suffers from inter-domain gaps; self-training yields biased pseudo-labels dominated by common object types and existing methods address only single factors (size, shape, density).
Approach
Data-driven GMM-based grouping of object descriptors from label sets, with a group-equivariant spatial feature (cross-attention between group queries and BEV features) that gives each group equal detection attention, plus an explorative update of group parameters from pseudo-labels.
Key results
- On Waymo to KITTI with SECOND-IoU, the method reaches 86.94 APBEV / 73.70 AP3D, improving over ST3D by 4.75/11.88 and over DTS by 1.14/2.21.
- On Waymo to NuScenes with SECOND-IoU, the method reaches 43.84 APBEV / 24.42 AP3D, outperforming the best prior method DTS by 3.27/2.91 in APBEV/AP3D.
- Ablation on Waymo to NuScenes shows grouping alone raises APBEV from 35.92 to 39.48 and the full set of components reaches 44.47 APBEV.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GroupEXP-DA instead of DTS | SECOND-IOU · Waymo → KITTI | APBEV/AP3D: 85.80/71.50 -> 86.94/73.70 | Helped |
| Use GroupEXP-DA instead of GPA-3D | PointPillars · Waymo → KITTI | APBEV/AP3D: 77.29/50.84 -> 78.44/54.11 | Helped |
| Use GroupEXP-DA instead of DTS | SECOND-IOU · NuScenes → KITTI | APBEV/AP3D: 81.4/66.6 -> 81.47/68.2 | Helped |
| Use GroupEXP-DA instead of DTS | PointPillars · NuScenes → KITTI | APBEV/AP3D: 79.5/51.8 -> 81.89/52.84 | Helped |
| Use GroupEXP-DA instead of DTS | SECOND-IOU · Waymo → NuScenes | APBEV/AP3D: 41.2/23.0 -> 43.84/24.42 | Helped |
| Use GroupEXP-DA instead of DTS | PointPillars · Waymo → NuScenes | APBEV/AP3D: 42.2/21.5 -> 44.31/22.15 | Helped |
| Add grouping to the baseline (ablation) | SECOND-IOU · Waymo → NuScenes | APBEV/AP3D: 35.92/20.19 -> 39.48/22.66 | Helped |
| Add Latt to grouping (ablation) | GroupEXP-DA · Waymo → NuScenes | APBEV/AP3D: 39.48/22.66 -> 40.82/23.14 | Helped |
| Add Lrepel to grouping (ablation) | GroupEXP-DA · Waymo → NuScenes | APBEV/AP3D: 39.48/22.66 -> 41.14/23.61 | Helped |
| Add both Latt and Lrepel to grouping (ablation) | GroupEXP-DA · Waymo → NuScenes | APBEV/AP3D: 39.48/22.66 -> 41.64/24.55 | Helped |
| Add explorative update to grouping, Latt, and Lrepel (ablation) | GroupEXP-DA · Waymo → NuScenes | APBEV/AP3D: 41.64/24.55 -> 44.47/25.91 | Helped |
| Replace proximity-based grouping with GMM-based grouping (ablation) | GroupEXP-DA, α=0.4 · NuScenes → KITTI | APBEV/AP3D: 79.82/64.77 -> 81.37/66.95 | Helped |
Surprising findings
- Making groups distinctive (repel loss) contributed more than intra-group cohesion (attraction loss) in ablations.
- Performance decreases when the number of groups exceeds about 4, attributed to overfitting or underfitting.
Limitations
- Struggles to detect objects with extremely sparse foreground points.
- Performance is sensitive to the number of groups, with degradation beyond about 4 groups.
Experimental setup
Datasets: Waymo NuScenes KITTI
Benchmarks: KITTI NuScenes Waymo
Metrics: AP_BEV AP_3D closed gap
Baselines: Source Only SN 3D-CoCo ST3D ST3D++ GPA-3D DTS ReDB
Builds on: ST3D PointPillars SECOND-IoU GPA-3D
Closest papers by meaning
- LION: Linear Group RNN for 3D Object Detection in Point Clouds (NeurIPS 2024 · similarity 0.88)
- Unified Domain Generalization and Adaptation for Multi-View 3D Object Detection (NeurIPS 2024 · similarity 0.87)
- Towards Generalizable Multi-Camera 3D Object Detection via Perspective Rendering (AAAI 2025 · similarity 0.87)
- DI-V2X: Learning Domain-Invariant Representation for Vehicle-Infrastructure Collaborative 3D Object Detection (AAAI 2024 · similarity 0.87)
- One for All: Multi-Domain Joint Training for Point Cloud Based 3D Object Detection (NeurIPS 2024 · similarity 0.86)
- An Empirical Study of the Generalization Ability of Lidar 3D Object Detectors to Unseen Domains (CVPR 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.