Research map / Detection and segmentation / 3D detection and driving
Decoupled Pseudo-labeling for Semi-Supervised 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 Decoupled Pseudo-labeling (DPL) for semi-supervised monocular 3D object detection, pairing homography-based BEV pseudo-label mining for 3D attributes with depth gradient projection to remove conflicts from noisy depth pseudo-label gradients, built on MonoFlex.
Problem
Semi-supervised monocular 3D object detection suffers from pseudo-labels whose 2D confidence does not reflect 3D (depth/orientation) quality, and noisy depth pseudo-label supervision that conflicts with reliable supervision.
Approach
Decouple pseudo-label generation: 2D attributes filtered by classification confidence, while 3D attributes are mined via an iterative homography between image and BEV planes that measures localization deviation; plus project conflicting depth gradients onto the reliable-supervision gradient direction.
Key results
- On KITTI test (Car, AP3D|R40), DPL with MonoFlex reaches 24.19 / 16.67 / 13.83 (Easy/Mod/Hard), +4.10 Mod over the MonoFlex baseline and +2.33 over Mix-Teaching.
- On KITTI val with 10% labeled data, DPL improves MonoFlex Mod AP3D by +4.29 (8.19 vs 3.90).
- Ablation on KITTI val: DPG adds +1.21 Mod AP3D over the confidence-threshold baseline and DGP adds +0.62, reaching 19.85 Mod AP3D.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DPL to MonoFlex | MonoFlex · KITTI test, Car | AP3D|R40 Mod.: 12.57 -> 16.67 | Helped |
| Add DPL to MonoFlex | MonoFlex · KITTI test, Car | APBEV|R40 Mod.: 17.94 -> 22.12 | Helped |
| Add DPL to MonoFlex | MonoFlex · KITTI validation, Car, 10% labeled | AP3D|R40 Mod: 3.90 -> 8.19 | Helped |
| Increase unlabeled data from 5K to 35K | DPL with MonoFlex · KITTI validation, Car | AP3D|R40 Mod: 18.76 -> 19.84 | Helped |
| Add decoupled pseudo-label generation (DPG) (ablation) | MonoFlex SSM3OD baseline · KITTI validation, Car | AP3D|R40 Mod.: 18.02 -> 19.23 | Helped |
| Add depth gradient projection (DGP) (ablation) | MonoFlex with DPG · KITTI validation, Car | AP3D|R40 Mod.: 19.23 -> 19.85 | Helped |
| Use detection-distance filtering instead of classification-confidence filtering (ablation) | MonoFlex SSM3OD · KITTI validation, Car | AP3D|R40 Mod: 18.02 -> 18.52 | Helped |
| Use homography-mined pseudo-labels without decoupling 2D and 3D supervision (ablation) | MonoFlex SSM3OD · KITTI validation, Car | AP3D|R40 Mod: 18.02 -> 19.04 | Helped |
| Decouple 2D and 3D pseudo-label generation (ablation) | MonoFlex with homography mining · KITTI validation, Car | AP3D|R40 Mod: 19.04 -> 19.23 | Helped |
| Apply DPL | MonoDLE · KITTI validation, Car | AP3D|R40 Mod: 13.87 -> 15.67 | Helped |
| Apply DPL | PGD · KITTI validation, Car | AP3D|R40 Mod: 13.22 -> 15.34 | Helped |
Surprising findings
- Classification confidence correlates weakly with depth prediction quality (PCC -0.196), so high-confidence pseudo-labels often have poor depth.
- Depth pseudo-label gradients conflict more often with reliable supervision than reliable supervision conflicts among itself.
Limitations
- Gains on the hard category are smaller than on easy/moderate categories, and distance-based filtering alone can discard useful far-away pseudo-labels.
- Method relies on flat-ground homography assumption and an estimated homography from teacher predictions.
- Evaluated only on KITTI Car category with MonoFlex as the main detector.
Experimental setup
Datasets: KITTI KITTI raw unlabeled video sequences
Benchmarks: KITTI test KITTI validation
Metrics: AP3D|R40 APBEV|R40
Baselines: Mix-Teaching MVC-MonoDet MonoFlex MonoDLE PGD GUPNet MonoDETR CaDDN MonoDTR
Builds on: MonoFlex Mix-Teaching Laplacian aleatoric uncertainty loss mean teacher (EMA teacher-student)
Closest papers by meaning
- Leveraging Temporal Cues for Semi-Supervised Multi-View 3D Object Detection (CVPR 2025 · similarity 0.89)
- Dual-Perspective Knowledge Enrichment for Semi-supervised 3D Object Detection (AAAI 2024 · similarity 0.88)
- FD3D: Exploiting Foreground Depth Map for Feature-Supervised Monocular 3D Object Detection (AAAI 2024 · similarity 0.87)
- Fully Data-Driven Pseudo Label Estimation for Pointly-Supervised Panoptic Segmentation (AAAI 2024 · similarity 0.86)
- LabelAny3D: Label Any Object 3D in the Wild (NeurIPS 2025 · similarity 0.86)
- Semi-supervised 3D Object Detection with PatchTeacher and PillarMix (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.