Research map / Detection and segmentation / 3D detection and driving
Geometry-Guided Domain Generalization for Monocular 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 MonoGDG, a geometry-guided domain generalization framework for monocular 3D object detection that combines camera-level reprojection (intrinsic unification, orientation randomization, FOV unification) with geometry-dependent feature disentanglement built on FCOS3D.
Problem
Monocular 3D object detectors degrade under real-world domain gaps (camera intrinsics, extrinsics, image appearance), and conventional domain-invariant feature learning causes negative transfer because geometry targets depend on the domain.
Approach
Reproject images to a unified spherical meta-camera so intrinsics are unified and orientation is randomized via image rotation/translation, and disentangle features into domain-shared and domain-specific branches with a depth-binned domain discriminator and domain-aware geometry regression attention, so geometry features are not destroyed by domain invariance.
Key results
- On nuScenes→KITTI, MonoGDG reaches 32.39 AP11 moderate 3D-BEV-style BEV AP versus 28.99 for DGMono3D (the best prior baseline).
- On PreSIL+nuScenes→KITTI, AP40 3D moderate improves to 23.31 versus 20.08 for DGMono3D.
- Ablation shows intrinsic parameter unification improves AP40 BEV moderate from 5.27 to 16.94 range-wise over the baseline without reprojection; full model reaches 22.01 AP40 3D moderate.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add intrinsic parameter unification (ablation) | MonoGDG camera reprojection · nuScenes→KITTI | AP40 3D, Moderate: 5.27 -> 16.94 | Helped |
| add FOV range unification to IPU (ablation) | MonoGDG camera reprojection · nuScenes→KITTI | AP40 3D, Moderate: 16.94 -> 19.48 | Helped |
| add camera orientation randomization to IPU (ablation) | MonoGDG camera reprojection · nuScenes→KITTI | AP40 3D, Moderate: 16.94 -> 21.62 | Helped |
| add IPU, COR, and FOVRU (ablation) | MonoGDG camera reprojection · nuScenes→KITTI | AP40 3D, Moderate: 5.27 -> 22.01 | Helped |
| use domain invariant learning (ablation) | MonoGDG feature learning · nuScenes→KITTI | AP40 3D, Moderate: 15.82 -> 10.41 | Hurt |
| disentangle features into domain-specific and domain-shared branches (ablation) | MonoGDG feature learning · nuScenes→KITTI | AP40 3D, Moderate: 15.82 -> 16.36 | Helped |
| add domain-aware geometry regression attention (ablation) | MonoGDG feature disentanglement · nuScenes→KITTI | AP40 3D, Moderate: 16.36 -> 20.73 | Helped |
| add depth-disentangled domain discriminator (ablation) | MonoGDG feature disentanglement · nuScenes→KITTI | AP40 3D, Moderate: 16.36 -> 20.59 | Helped |
| add both DAGRA and DDDD (ablation) | MonoGDG feature disentanglement · nuScenes→KITTI | AP40 3D, Moderate: 15.82 -> 22.01 | Helped |
| add image appearance domain discriminator (ablation) | MonoGDG · PreSIL+nuScenes→KITTI | AP40 3D, Moderate: 21.94 -> 23.31 | Helped |
| use intrinsic parameter unification instead of GAMS (ablation) | MonoGDG · nuScenes→KITTI | AP40 3D, Moderate: 19.84 -> 22.01 | Helped |
| use MonoGDG instead of DGMono3D | FCOS3D · nuScenes→KITTI | AP40 3D, Moderate: 18.55 -> 22.01 | Helped |
Surprising findings
- Conventional domain adversarial learning causes a substantial performance drop for M3OD (negative transfer), with accuracy decreasing as domain invariance increases.
- The method surpasses the Oracle (fully supervised on target) in most settings.
Limitations
- Camera position (extrinsic translation) randomization is not addressed because perspective transformation needs pixel-wise depth.
- Evaluation limited to driving datasets with FCOS3D as the sole detector.
Experimental setup
Datasets: nuScenes Lyft PreSIL KITTI
Benchmarks: KITTI nuScenes PreSIL
Metrics: AP11 AP40 BEV AP 3D AP AP ATE ASE AOE
Baselines: Source Only Oracle STMono3D DGMono3D GAMS
Builds on: FCOS3D STMono3D DGMono3D domain adversarial training
Closest papers by meaning
- Unified Domain Generalization and Adaptation for Multi-View 3D Object Detection (NeurIPS 2024 · similarity 0.89)
- MonoDGP: Monocular 3D Object Detection with Decoupled-Query and Geometry-Error Priors (CVPR 2025 · similarity 0.89)
- Generalizable Multi-Camera 3D Object Detection from a Single Source via Fourier Cross-View Learning (ICML 2025 · similarity 0.87)
- Training an Open-Vocabulary Monocular 3D Detection Model without 3D Data (NeurIPS 2024 · similarity 0.87)
- MonoTAKD: Teaching Assistant Knowledge Distillation for Monocular 3D Object Detection (CVPR 2025 · similarity 0.87)
- UniMODE: Unified Monocular 3D 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.