Research map / 3D vision / 3D generation
MonoTAKD: Teaching Assistant Knowledge Distillation for 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 MonoTAKD, a teaching-assistant knowledge distillation framework for monocular 3D detection that distills from a LiDAR teacher via a camera-based TA with intra-modal and cross-modal residual distillation.
Problem
Monocular 3D object detection suffers from depth ambiguity, and direct cross-modal distillation from a LiDAR teacher to a camera student is ineffective due to a large feature representation gap.
Approach
Insert a camera-based teaching assistant that uses GT depth to produce optimal BEV features (distilled intra-modally to the student), and distill the residual between LiDAR teacher and TA BEV features as 3D spatial cues, plus a spatial alignment module and feature fusion module.
Key results
- On KITTI test Car, MonoTAKD reaches 27.91 / 19.43 / 16.51 AP3D (easy/moderate/hard), beating the second-best method by +2.30, +2.06, +1.21.
- On nuScenes val, TAKD with BEVFormer R101 reaches 0.558 NDS and 0.451 mAP, and with BEVDepth R101 reaches 0.564 NDS and 0.466 mAP.
- Ablation on KITTI val: MonoTAKD improves AP3D over the no-distillation CaDDN baseline by +10.01 (easy), +6.45 (moderate), +6.44 (hard).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add intra-modal distillation to CaDDN (ablation) | CaDDN student · KITTI val, Car | AP3D: 24.35 -> 31.11; 16.16 -> 20.24; 13.44 -> 16.91 | Helped |
| add direct cross-modal distillation instead of no feature distillation (ablation) | CaDDN student · KITTI val, Car | AP3D: 24.35 -> 28.22; 16.16 -> 18.29; 13.44 -> 15.10 | Helped |
| replace direct cross-modal distillation with cross-modal residual distillation (ablation) | CaDDN student · KITTI val, Car | AP3D: 28.22 -> 29.68; 18.29 -> 19.57; 15.10 -> 16.22 | Helped |
| add cross-modal residual distillation to intra-modal distillation (ablation) | CaDDN student · KITTI val, Car | AP3D: 31.11 -> 34.36; 20.24 -> 22.61; 16.91 -> 19.88 | Helped |
| add ASPP to the spatial alignment module (ablation) | MonoTAKD · KITTI val, Car | AP3D: 31.36 -> 32.28; 20.80 -> 21.45; 16.53 -> 17.18 | Helped |
| add deformable convolution to the spatial alignment module (ablation) | MonoTAKD · KITTI val, Car | AP3D: 31.36 -> 32.46; 20.80 -> 21.51; 16.53 -> 17.99 | Helped |
| add ASPP and deformable convolution (ablation) | MonoTAKD · KITTI val, Car | AP3D: 31.36 -> 33.19; 20.80 -> 21.81; 16.53 -> 18.23 | Helped |
| add SENet to the spatial alignment module (ablation) | MonoTAKD · KITTI val, Car | AP3D: 33.19 -> 33.26; 21.81 -> 21.47; 18.23 -> 18.61 | Mixed |
| add the feature fusion module (ablation) | MonoTAKD · KITTI val, Car | AP3D: 33.26 -> 34.36; 21.47 -> 22.61; 18.61 -> 19.88 | Helped |
| use a camera-based TA model matching the student architecture rather than a heterogeneous TA (ablation) | MonoTAKD · KITTI val, Car | AP3D: 30.74 -> 33.18; 20.35 -> 21.97; 17.69 -> 18.55 | Mixed |
Surprising findings
- Distilling across heterogeneous architectures (transformer to CNN) may be ineffective due to substantial differences in feature encoding.
- Learning residual features (LiDAR-minus-TA BEV differences) outperforms replicating the full LiDAR BEV features.
Limitations
- SAM is optional and adds some FLOPs overhead.
- Fair comparison is hindered by heterogeneous experimental settings and architectures, as the authors note.
- The TA uses ground-truth depth maps during training, which is not available at test time.
Experimental setup
Datasets: KITTI3D nuScenes KITTI raw
Benchmarks: KITTI3D nuScenes
Metrics: AP3D APBEV NDS mAP
Baselines: CMKD ADD MonoDistill MonoDETR CaDDN MonoCD OccupancyM3D DistillBEV STXD BEVDistill LabelDistill BEVDepth
Builds on: CMKD CaDDN MonoDETR BEVDistill BEVDepth
Closest papers by meaning
- FD3D: Exploiting Foreground Depth Map for Feature-Supervised Monocular 3D Object Detection (AAAI 2024 · similarity 0.88)
- MonoDiff: Monocular 3D Object Detection and Pose Estimation with Diffusion Models (CVPR 2024 · similarity 0.88)
- MonoDGP: Monocular 3D Object Detection with Decoupled-Query and Geometry-Error Priors (CVPR 2025 · similarity 0.88)
- MonoMAE: Enhancing Monocular 3D Detection through Depth-Aware Masked Autoencoders (NeurIPS 2024 · similarity 0.88)
- Training an Open-Vocabulary Monocular 3D Detection Model without 3D Data (NeurIPS 2024 · similarity 0.88)
- MonoPlace3D: Learning 3D-Aware Object Placement for 3D Monocular Detection (CVPR 2025 · 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.