Research map / Detection and segmentation / 3D detection and driving
FD3D: Exploiting Foreground Depth Map for Feature-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 FD3D, a feature-supervised monocular 3D detector that trains an instructive feature generation network on image plus annotation foreground object-wise depth maps (AFOD), then distills its features into a monocular MonoDLE/DD3D student via pixel-level and adversarial distribution-level supervision.
Problem
Monocular 3D object detection lacks spatial depth cues; existing distillation from LiDAR- or stereo-based teachers suffers from modality/domain gaps and expensive extra sensors.
Approach
Replace multi-modal teacher with a monocular-input generation network that uses ground-truth object-wise depth maps (AFOD) as a spatial indicator, via a vision-depth cross-attention association module, to produce instructive features; supervise the student with channel-wise projection MSE loss plus an adversarial scoring block. No extra inference cost.
Key results
- On KITTI test Car, FD3D reaches 25.38 / 17.12 / 14.50 AP3D (easy/mod/hard), improving on the best prior no-extra-data result by +0.69/+0.66/+0.55.
- On nuScenes test, FD3D reaches 0.431 mAP and 0.485 NDS, +1.3% mAP and +0.8% NDS over the DD3D base model.
- Ablation on KITTI val shows the auxiliary feature supervision (CPL + ASB) raises AP3D moderate from 13.66 to 20.23 over the MonoDLE-style baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add FD3D auxiliary feature supervision to DD3D | DD3D · nuScenes test | mAP: 0.418 -> 0.431 | Helped |
| Add FD3D auxiliary feature supervision to DD3D | DD3D · nuScenes test | NDS: 0.477 -> 0.485 | Helped |
| Add FD3D to MonoDLE | MonoDLE · KITTI test, Car moderate | AP3D: 12.26 -> 17.12 | Helped |
| Add FD3D to MonoDLE | MonoDLE · KITTI test, Car moderate | APBEV: 18.89 -> 23.72 | Helped |
| Use cross-attention instead of direct concatenation-plus-MLP association (ablation) | FD3D IFGN · KITTI validation, Car | AP3D@IoU=0.7 moderate: 57.32 -> 57.45 | Helped |
| Add preset 3D coordinates to cross-attention association (ablation) | FD3D IFGN · KITTI validation, Car | AP3D@IoU=0.7 moderate: 57.45 -> 58.54 | Helped |
| Weight coordinates with visual features (ablation) | FD3D IFGN · KITTI validation, Car | AP3D@IoU=0.7 moderate: 58.54 -> 63.52 | Helped |
| Weight coordinates with depth knowledge (ablation) | FD3D IFGN · KITTI validation, Car | AP3D@IoU=0.7 moderate: 58.54 -> 62.97 | Helped |
| Weight coordinates with cross-attention features (ablation) | FD3D IFGN · KITTI validation, Car | AP3D@IoU=0.7 moderate: 58.54 -> 67.10 | Helped |
| Add channel-wise projection and adversarial scoring supervision (ablation) | MonoDLE · KITTI validation, Car | AP3D@IoU=0.7 moderate: 13.66 -> 19.43 | Helped |
| Add channel-wise projection layer only (ablation) | MonoDLE · KITTI validation, Car | AP3D@IoU=0.7 moderate: 13.66 -> 18.16 | Helped |
| Add full pixel- and distribution-level auxiliary supervision (ablation) | MonoDLE · KITTI validation, Car | AP3D@IoU=0.7 moderate: 13.66 -> 20.23 | Helped |
Surprising findings
- AFOD-based instructive features outperform full-range LiDAR depth as guidance, which the authors attribute to background noise such as buildings corrupting LiDAR depth.
Limitations
- Relies on ground-truth object annotations (AFOD) during training, so the instructive feature generator depends on label quality.
- Gains over the base models are modest on nuScenes (+1.3% mAP).
Experimental setup
Datasets: KITTI nuScenes
Benchmarks: KITTI nuScenes
Metrics: AP3D APBEV mAP NDS mATE mAOE
Baselines: MonoDTR MonoDistill DCD SGM3D OPA-3D NeurOCS PGD MonoDLE MonoEF GUPNet HomoLoss MonoJSG
Builds on: MonoDLE DD3D MonoDistill MLP-Mixer
Closest papers by meaning
- Weakly Supervised Monocular 3D Detection with a Single-View Image (CVPR 2024 · similarity 0.89)
- MonoMAE: Enhancing Monocular 3D Detection through Depth-Aware Masked Autoencoders (NeurIPS 2024 · similarity 0.89)
- MonoDGP: Monocular 3D Object Detection with Decoupled-Query and Geometry-Error Priors (CVPR 2025 · similarity 0.88)
- MonoCD: Monocular 3D Object Detection with Complementary Depths (CVPR 2024 · similarity 0.88)
- MonoTAKD: Teaching Assistant Knowledge Distillation for Monocular 3D Object Detection (CVPR 2025 · similarity 0.88)
- Training an Open-Vocabulary Monocular 3D Detection Model without 3D Data (NeurIPS 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.