Research map / Detection and segmentation / 3D detection and driving
CMDA: Cross-Modal and Domain Adversarial Adaptation for LiDAR-Based 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 CMDA, an unsupervised domain adaptation method for LiDAR 3D object detection that aligns LiDAR BEV features with camera BEV features (CMKI) and adds adversarial domain-discriminative self-training with cross-domain point mix-up (CDAN).
Problem
LiDAR-based 3D object detectors degrade substantially when deployed on target domains with different point density, weather, or geography, and existing LiDAR-only UDA methods struggle to learn domain-agnostic semantic cues.
Approach
Uses camera-derived BEV features as a semantic bridge: a LiDAR encoder is pretrained to match image BEV features via an L2 loss, then self-training with cross-domain sector mix-up, a gradient-reversal domain discriminator on instance features, and BEV grid-wise entropy regularization adapts the detector to the target domain.
Key results
- On Waymo to nuScenes with PV-RCNN, CMDA reaches 44.44 BEV AP / 26.41 3D AP, versus 34.50 / 21.47 for direct transfer.
- On nuScenes to KITTI with SECOND-IoU, CMDA reaches 82.13 BEV AP / 68.95 3D AP, versus 75.94 / 54.13 for ST3D.
- Ablation on nuScenes to KITTI shows CMKI and CDAN each add gains over ST3D alone (e.g., SECOND-IoU 3D AP 54.13 to 60.04 with CMKI and 68.95 with CDAN).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CMDA instead of ST3D | SECOND-IoU · nuScenes → Waymo | BEV AP / 3D AP: 45.35 / 27.12 -> 46.79 / 29.42 | Helped |
| use CMDA instead of ST3D | PV-RCNN · nuScenes → Waymo | BEV AP / 3D AP: 52.50 / 36.21 -> 58.57 / 45.58 | Helped |
| use CMDA instead of DTS | SECOND-IoU · nuScenes → KITTI | BEV AP / 3D AP: 81.40 / 66.60 -> 82.13 / 68.95 | Helped |
| use CMDA instead of DTS | PV-RCNN · nuScenes → KITTI | BEV AP / 3D AP: 83.90 / 71.80 -> 84.85 / 75.02 | Helped |
| use CMDA w/ LD instead of DTS | SECOND-IoU · Waymo → nuScenes | BEV AP / 3D AP: 41.20 / 23.00 -> 42.81 / 24.64 | Helped |
| use CMDA w/ LD instead of DTS | PV-RCNN · Waymo → nuScenes | BEV AP / 3D AP: 44.00 / 26.20 -> 44.44 / 26.41 | Helped |
| add CMKI (ablation) | Direct Transfer · nuScenes → Waymo | BEV AP / 3D AP: 39.18 / 20.78 -> 44.41 / 22.26 | Helped |
| add CMKI (ablation) | Direct Transfer · nuScenes → Waymo | BEV AP / 3D AP: 41.30 / 25.89 -> 47.28 / 28.57 | Helped |
| add CMKI (ablation) | Direct Transfer · nuScenes → KITTI | BEV AP / 3D AP: 51.84 / 17.92 -> 63.86 / 37.22 | Helped |
| add CMKI (ablation) | Direct Transfer · nuScenes → KITTI | BEV AP / 3D AP: 68.15 / 37.17 -> 72.12 / 40.17 | Helped |
| add CDAN to ST3D + CMKI (ablation) | SECOND-IoU · nuScenes → Waymo | BEV AP / 3D AP: 45.43 / 28.63 -> 46.79 / 29.42 | Helped |
| add CDAN to ST3D + CMKI (ablation) | PV-RCNN · nuScenes → Waymo | BEV AP / 3D AP: 53.04 / 42.75 -> 58.57 / 45.58 | Helped |
Limitations
- Requires multi-modal camera data during source training, limiting use where images are unavailable.
- Gains are reported mainly on car category and depend on the chosen baseline detectors.
Experimental setup
Datasets: nuScenes Waymo KITTI
Benchmarks: nuScenes to Waymo nuScenes to KITTI Waymo to nuScenes
Metrics: BEV AP 3D AP Closed Gap
Baselines: Direct Transfer SN ST3D ST3D++ LiDAR Distillation DTS 3D-CoCo contrastive learning
Builds on: ST3D Lift-Splat-Shoot SECOND PV-RCNN gradient reversal layer
Closest papers by meaning
- Multi-Modal Assistance for Unsupervised Domain Adaptation on Point Cloud 3D Object Detection (AAAI 2026 · similarity 0.90)
- CL3D: Unsupervised Domain Adaptation for Cross-LiDAR 3D Detection (AAAI 2023 · similarity 0.90)
- An Empirical Study of the Generalization Ability of Lidar 3D Object Detectors to Unseen Domains (CVPR 2024 · similarity 0.88)
- Unified Domain Generalization and Adaptation for Multi-View 3D Object Detection (NeurIPS 2024 · similarity 0.87)
- MOS: Model Synergy for Test-Time Adaptation on LiDAR-Based 3D Object Detection (ICLR 2025 · similarity 0.87)
- Geometry-Guided Domain Generalization for Monocular 3D Object Detection (AAAI 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.