Research map / Detection and segmentation / 3D detection and driving
MonoCD: Monocular 3D Object Detection with Complementary Depths
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 MonoCD, a CenterNet-based monocular 3D detector that adds a global-clue "complementary depth" branch (from a predicted ground-plane horizon) and geometric complementary form so multiple depth errors have opposite signs, improving KITTI accuracy.
Problem
Multi-depth ensembles in monocular 3D object detection produce depth errors with the same sign (coupling), so they cannot cancel each other and the combined depth accuracy is limited.
Approach
Add a depth branch built from a global clue (object elevation y derived from a predicted ground-plane horizon shared across the image) plus a geometric formulation complementary in form to the keypoint depth, so that errors have opposite signs and partially cancel under uncertainty-weighted averaging.
Key results
- On KITTI test Car moderate, MonoCD reaches 16.59 AP3D and 22.81 APBEV, with 36 ms runtime, without extra data.
- Plugged into MonoDLE, MonoFlex, and MonoCon on KITTI val, AP3D moderate improves by 1.43, 0.76, and 0.36 points respectively.
- Flipping the direct-depth branch on the MonoFlex baseline shows AP40 rising from 17.51 to 36.12 as the flipped proportion goes from 0% to 100%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add complementary depth to MonoFlex | MonoFlex · KITTI val Car | AP3D Moderate: 17.51 -> 18.27 | Helped |
| add complementary depth to MonoDLE | MonoDLE · KITTI val Car | AP3D Moderate: 13.66 -> 15.09 | Helped |
| add complementary depth to MonoCon | MonoCon · KITTI val Car | AP3D Moderate: 19.01 -> 19.37 | Helped |
| predict y from global rather than local clues (ablation) | MonoFlex · KITTI val Car | AP3D Moderate: 13.49 -> 17.85 | Helped |
| use a fixed 1.65-meter y source instead of the baseline source (ablation) | MonoFlex · KITTI val Car | AP3D Moderate: 17.51 -> 15.86 | Hurt |
| use the ground-truth ground-plane equation for y (ablation) | MonoFlex · KITTI val Car | AP3D Moderate: 17.51 -> 19.43 | Helped |
| use edge detection to obtain the horizon slope (ablation) | MonoFlex global-clue branch · KITTI val Car | AP3D Moderate: 17.85 -> 15.97 | Hurt |
| use dilated convolution for the global-clue branch (ablation) | MonoFlex global-clue branch · KITTI val Car | AP3D Moderate: 17.85 -> 18.27 | Helped |
| use complementary depth form Eq. (14) instead of Eq. (12) (ablation) | MonoFlex · KITTI val Car | AP3D Moderate: 17.62 -> 18.27 | Helped |
| use complementary depth form Eq. (16) instead of Eq. (14) (ablation) | MonoFlex · KITTI val Car | AP3D Moderate: 18.27 -> 15.97 | Hurt |
Surprising findings
- About 95% of existing multi-depth predictions share the same error sign on KITTI, which limits ensemble gains.
- A complementary depth form with higher error-sign opposition (Eq. 16) performed worse overall because its depth estimate was unstable.
Limitations
- Performance is limited by the accuracy of the vertical position of objects.
- Complementary effect may be lost when the ground plane is undulating.
Experimental setup
Datasets: KITTI 3D Object
Benchmarks: KITTI 3D Object test set KITTI validation set
Metrics: AP3D APBEV Complementarity Score MAE ESOP
Baselines: DDMP-3D Kinematic3D AutoShape DCD MonoRUn CaDDN MonoDTR SMOKE MonoDLE MonoRCNN MonoFlex MonoGround
Builds on: CenterNet MonoFlex MonoDLE MonoCon GPENet
Closest papers by meaning
- Objects Are Different: Flexible Monocular 3D Object Detection (CVPR 2021 · similarity 0.92)
- MonoDGP: Monocular 3D Object Detection with Decoupled-Query and Geometry-Error Priors (CVPR 2025 · similarity 0.91)
- MonoMAE: Enhancing Monocular 3D Detection through Depth-Aware Masked Autoencoders (NeurIPS 2024 · similarity 0.90)
- MonoDiff: Monocular 3D Object Detection and Pose Estimation with Diffusion Models (CVPR 2024 · similarity 0.89)
- RefineAny3D: Depth Refinement as Semantic Alignment for Monocular 3D Detection (NeurIPS 2026 · similarity 0.88)
- FD3D: Exploiting Foreground Depth Map for Feature-Supervised Monocular 3D Object Detection (AAAI 2024 · similarity 0.88)
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.