Research map / Detection and segmentation / 3D detection and driving
MonoDGP: Monocular 3D Object Detection with Decoupled-Query and Geometry-Error Priors
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 MonoDGP, a DETR-style monocular 3D detector that predicts perspective-invariant geometry (depth) errors to correct geometric depth, decouples a 2D visual decoder for query initialization, and adds a Region Segmentation Head for enhanced features and segment embeddings.
Problem
Monocular 3D object detection relies on perspective projection for geometric depth, but the 2D bounding-box height often fails to represent the actual central height, causing depth errors; multi-depth prediction is complex and under-exploits geometric depth.
Approach
Replace direct depth prediction with a perspective-invariant depth error added to geometric depth computed from the 2D box height, combined with a decoupled 2D visual decoder that initializes object queries and a U-Net-style region segmentation head that enhances foreground features and supplies BERT-like segment embeddings.
Key results
- On KITTI val, moderate AP3D reaches 22.34% vs. 20.22% for the baseline (+2.12), with the full model improving moderate AP3D by 1.73% over MonoDETR on val.
- Depth error prediction (GD + Depth Error) achieves 22.34% moderate AP3D, outperforming weighted fusion of multiple depths (21.64%) and bounding-box height error (21.18%).
- Achieves state-of-the-art monocular results on KITTI test and val for car category without extra data (except one easy APBEV metric where OccupancyM3D, using LiDAR data, is higher).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace MonoDDE with MonoDGP | KITTI test, Car Moderate · KITTI test, Car Moderate | AP3D|R40: 17.14 -> 18.72 | Helped |
| replace MonoDETR with MonoDGP | KITTI val, Car Moderate · KITTI val, Car Moderate | AP3D|R40: 20.61 -> 22.34 | Helped |
| add decoupled query (ablation) | MonoDGP · KITTI val, Car Moderate | AP3D: 21.48 -> 22.34 | Helped |
| add RSH (ablation) | MonoDGP · KITTI val, Car Moderate | AP3D: 21.93 -> 22.34 | Helped |
| set segment embedding threshold to 0.3 instead of 0.5 (ablation) | MonoDGP · KITTI val, Car Moderate | AP3D|R40: 21.64 -> 21.51 | Hurt |
| set segment embedding threshold to 0.7 instead of 0.5 (ablation) | MonoDGP · KITTI val, Car Moderate | AP3D|R40: 21.64 -> 21.54 | Hurt |
| remove segment embeddings (ablation) | MonoDGP · KITTI val, Car Moderate | AP3D|R40: 21.43 vs 21.64 | Hurt |
| use direct depth prediction instead of the full method (ablation) | MonoDGP · KITTI val, Car Moderate | AP3D|R40: 19.23 vs 22.34 | Hurt |
| use geometric depth from hbbox without error correction (ablation) | MonoDGP · KITTI val, Car Moderate | AP3D|R40: 7.99 vs 22.34 | Hurt |
| use geometric depth from hc instead of the proposed correction (ablation) | MonoDGP · KITTI val, Car Moderate | AP3D|R40: 20.73 vs 22.34 | Hurt |
| fuse depth map, direct depth, and geometric depth instead of using geometric depth error (ablation) | MonoDGP · KITTI val, Car Moderate | AP3D|R40: 21.64 vs 22.34 | Hurt |
| correct geometric depth with bounding box height error instead of depth error (ablation) | MonoDGP · KITTI val, Car Moderate | AP3D|R40: 21.18 vs 22.34 | Hurt |
Surprising findings
- Geometric depth computed from the 2D bounding-box height alone performs very poorly (7.99% moderate AP3D on val), far below geometric depth using the true central height h_c.
- Predicting bounding-box height error performs worse than predicting depth error, attributed to nonlinear gradient backpropagation through the denominator.
Limitations
- Gains over the MonoDETR baseline are modest (about 1-2 AP points), and the easy APBEV test result trails OccupancyM3D, which uses extra LiDAR data.
- Geometric analysis relies on assumptions of zero roll and pitch and negligible ground slope.
- Evaluation is limited to KITTI car category in the main tables.
Experimental setup
Datasets: KITTI
Benchmarks: KITTI
Metrics: AP3D APBEV AP40
Baselines: MonoDETR MonoCD MonoDDE GUPNet DID-M3D OccupancyM3D MonoDTR CaDDN FD3D
Builds on: MonoDETR DETR Deformable DETR BERT Mixup3D
Closest papers by meaning
- Rethinking Geometric Depth in Monocular 3D Object Detection: A Projection-Consistent Reformulation (NeurIPS 2026 · similarity 0.94)
- MonoCD: Monocular 3D Object Detection with Complementary Depths (CVPR 2024 · similarity 0.91)
- MonoRUn: Monocular 3D Object Detection by Reconstruction and Uncertainty Propagation (CVPR 2021 · similarity 0.91)
- MonoDiff: Monocular 3D Object Detection and Pose Estimation with Diffusion Models (CVPR 2024 · similarity 0.91)
- MonoGround: Detecting Monocular 3D Objects From the Ground (CVPR 2022 · similarity 0.91)
- MonoMAE: Enhancing Monocular 3D Detection through Depth-Aware Masked Autoencoders (NeurIPS 2024 · similarity 0.90)
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.