Research map / Detection and segmentation / 3D detection and driving
UniMODE: Unified 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.
Builds UniMODE, a BEV-based monocular 3D detector that unifies indoor and outdoor detection on Omni3D via a two-stage design, uneven BEV grid, sparse feature projection, and domain-aligned normalization and loss.
Problem
Unified monocular 3D object detection across indoor and outdoor scenes is hard because of diverse geometry properties and heterogeneous domain distributions, and existing BEV detectors converge unstably when trained on mixed data.
Approach
Adapts the BEV detection paradigm for unified monocular 3D detection with a CenterNet-style proposal stage that initializes queries, an uneven BEV depth grid that is fine near the camera and coarse far away, sparse BEV feature projection that drops low-confidence depth points, and domain adaptive layer normalization plus a class alignment loss to handle domain and label-space conflicts.
Key results
- On Omni3D, UniMODE with ConvNext-Base reaches 28.2% AP3D versus 23.3% for Cube RCNN; with DLA34 it reaches 25.5% AP3D (+2.2 points).
- The proposal head gives the largest ablation gain, adding 3.6% AP3D.
- Sparse BEV feature projection removes 82.6% of projection points at tau=1e-3 with no loss in AP3D (17.4% vs 17.4%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use UniMODE instead of Cube RCNN | Omni3D unified detector · OMNI3D | AP3D: 23.3% -> 28.2% | Helped |
| add the proposal head (ablation) | UniMODE · Omni3D unified detection | AP3D: 12.3% -> 15.9% | Helped |
| add the uneven BEV grid to the proposal-head model (ablation) | UniMODE · Omni3D unified detection | AP3D: 15.9% -> 16.6% | Helped |
| add sparse BEV feature projection to the proposal-head model (ablation) | UniMODE · Omni3D unified detection | AP3D: 15.9% -> 16.6% | No clear effect |
| add unified domain alignment to the model with proposal head, uneven grid, and sparse projection (ablation) | UniMODE · Omni3D unified detection | AP3D: 16.6% -> 17.4% | Helped |
| use uneven rather than even depth bins (ablation) | UniMODE · Omni3D unified detection | AP3D: 17.4% -> 15.3% | Hurt |
| reduce BEV grid size from 1 m to 0.5 m (ablation) | UniMODE · Omni3D unified detection | AP3D: 17.4% -> 18.1% | Helped |
| increase BEV grid size from 1 m to 2 m (ablation) | UniMODE · Omni3D unified detection | AP3D: 17.4% -> 16.5% | Hurt |
| set sparse projection threshold τ to 1e-3 instead of 0 (ablation) | UniMODE · Omni3D unified detection | AP3D: 17.4% -> 17.4% | No clear effect |
| set sparse projection threshold τ to 1e-2 instead of 0 (ablation) | UniMODE · Omni3D unified detection | AP3D: 17.4% -> 15.0% | Hurt |
| set sparse projection threshold τ to 1e-1 instead of 0 (ablation) | UniMODE · Omni3D unified detection | AP3D: 17.4% -> 4.7% | Hurt |
| use direct regression rather than no domain-adaptive strategy (ablation) | UniMODE · SUN-RGBD (out-of-domain) | AP3D: 12.3% -> 12.1% | Hurt |
Surprising findings
- Uneven depth binning (following the Eq. 1 formula) hurt performance (AP3D 15.3% vs 17.4%), contrary to the expectation that it would help.
- Cross-domain zero-shot transfer is weak, especially to Hypersim, with the authors noting that monocular depth estimation is ill-posed.
Limitations
- Zero-shot generalization to unseen data scenarios is still limited.
- Uneven depth binning did not improve results in the ablation.
Experimental setup
Datasets: Omni3D KITTI nuScenes SUN-RGBD ARKitScenes Objectron Hypersim
Benchmarks: Omni3D
Metrics: AP3D AP25 AP50
Baselines: Cube RCNN BEVFormer PETR M3D-RPN SMOKE FCOS3D PGD GUPNet ImVoxelNet
Builds on: Cube RCNN Deformable DETR CenterNet CaDDN layer normalization
Closest papers by meaning
- MonoDGP: Monocular 3D Object Detection with Decoupled-Query and Geometry-Error Priors (CVPR 2025 · similarity 0.89)
- Towards Generalizable Multi-Camera 3D Object Detection via Perspective Rendering (AAAI 2025 · similarity 0.88)
- Instance-Aware Multi-Camera 3D Object Detection with Structural Priors Mining and Self-Boosting Learning (AAAI 2024 · similarity 0.88)
- GeoBEV: Learning Geometric BEV Representation for Multi-view 3D Object Detection (AAAI 2025 · similarity 0.88)
- SeaBird: Segmentation in Bird's View with Dice Loss Improves Monocular 3D Detection of Large Objects (CVPR 2024 · similarity 0.88)
- Training an Open-Vocabulary Monocular 3D Detection Model without 3D Data (NeurIPS 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.