🚗 LeAD-M3D

Leveraging Asymmetric Distillation for Real-Time Monocular 3D Detection

Johannes Meier* · Jonathan Michel* · Oussema Dhaouadi · Yung-Hsu Yang · Christoph Reich · Zuria Bauer · Stefan Roth · Marc Pollefeys · Jacques Kaiser · Daniel Cremers

DeepScenario · ETH Zurich · TU Munich · University of Cambridge · MCML · TU Darmstadt · ELIZA · hessian.AI · Microsoft

---

LeAD-M3D predicts accurate 3D bounding boxes from a single RGB image in real time (KITTI val set).

---

✨ Highlights

Component	What it does
A2D2 — Asymmetric Augmentation Denoising Distillation	Transfers geometric depth knowledge from a clean-image teacher via a quality- and importance-weighted feature loss. No LiDAR needed.
CM-3D — 3D-aware Consistent Matching	Integrates 3D MGIoU into the assignment score for stable, precise supervision.
CGI-3D — Confidence-Gated 3D Inference	Routes only the top-k confident locations through the expensive 3D regression head — up to 3.6× faster than prior high-accuracy models.

---

📊 Results

KITTI Test — Car, AP40 @ IoU ≥ 0.7

Model	Params	GFLOPs	Time (FP32)	Time (TRT FP16)	Easy	Mod.	Hard
LeAD-M3D-N	3.8 M	14.2	8.2 ms	1.2 ms	24.31	16.49	14.14
LeAD-M3D-S	10.1 M	37.4	10.9 ms	2.1 ms	27.28	18.87	16.37
LeAD-M3D-M	19.7 M	87.6	12.9 ms	3.7 ms	28.08	19.47	17.66
LeAD-M3D-B	24.9 M	131.9	14.1 ms	4.5 ms	29.10	20.17	18.34
LeAD-M3D-X	36.3 M	217.1	24.3 ms	7.8 ms	30.76	21.20	18.76

Inference on a single image (384×1280) on NVIDIA Quadro RTX 8000, batch size 1. TRT FP16 uses TensorRT 8.6.1 with 16 GB workspace, median of 1000 runs.

KITTI Val — Car, AP40 3D (k=25)

Model	Params	GFLOPs	Time (FP32)	Time (TRT FP16)	Easy	Mod.	Hard
LeAD-M3D-N	3.8 M	14.2	8.2 ms	1.2 ms	23.85	18.38	16.69
LeAD-M3D-S	10.1 M	37.4	10.9 ms	2.1 ms	26.34	20.67	18.55
LeAD-M3D-M	19.7 M	87.6	12.9 ms	3.7 ms	27.98	22.40	19.76
LeAD-M3D-B	24.9 M	131.9	14.1 ms	4.5 ms	28.44	22.65	19.87
LeAD-M3D-X	36.3 M	217.1	24.3 ms	7.8 ms	28.30	22.96	21.09

---

⚙️ Installation

conda create -n lead_m3d python=3.9 -y
conda activate lead_m3d

# PyTorch — adjust for your CUDA version
pip install torch==2.0.1+cu117 torchvision==0.15.2+cu117 \
    --extra-index-url https://download.pytorch.org/whl/cu117

# Install this package in editable mode
pip install -e .

# Remaining dependencies
pip install -r requirements.txt

CPU-only or other CUDA versions: replace the torch line with the appropriate build from pytorch.org.

---

📁 Dataset Setup

Download the KITTI 3D Object Detection dataset and arrange it as:

kitti/
├── training/
│   ├── image_2/          # RGB images
│   ├── label_2/          # 3D annotations
│   └── calib/            # calibration files
├── testing/
│   ├── image_2/
│   └── calib/
└── ImageSets/
    ├── train.txt
    ├── val.txt
    └── test.txt

Set the dataset root in lead_m3d/cfg/datasets/kitti.yaml:

path: /path/to/kitti

---

📦 Model Zoo

Download pre-trained checkpoints and place them under checkpoints/:

Model	Val AP3D Mod.	Download
LeAD-M3D-N	18.41	lead_m3d_n.pt
LeAD-M3D-S	20.73	lead_m3d_s.pt
LeAD-M3D-M	22.46	lead_m3d_m.pt
LeAD-M3D-B	22.71	lead_m3d_b.pt
LeAD-M3D-X	23.01	lead_m3d_x.pt

---

🚀 Inference & Evaluation

python infer_kitti.py \
    --weights checkpoints/lead_m3d_b.pt \
    --data    lead_m3d/cfg/datasets/kitti.yaml \
    --topk    25

Argument	Default	Description
--config	—	Per-model config YAML (e.g. configs/kitti_b.yaml)
--weights	—	Path to .pt checkpoint
--data	lead_m3d/cfg/datasets/kitti.yaml	Dataset YAML
--topk	25	CGI-3D k: candidates per scale for the 3D head
--batch	8	Batch size
--output	runs/kitti_eval	Output directory
--half	False	FP16 inference
--device	""	CUDA device (0, 1, …, or cpu)

---

⚡ TensorRT Inference

pip install "onnx==1.14.1" onnxsim tensorrt==8.6.1
export LD_LIBRARY_PATH="$(python -c 'import os, torch; print(os.path.join(os.path.dirname(torch.__file__), "lib"))'):$LD_LIBRARY_PATH"

python export_trt.py --models n s m b x --device 0

Exports each model to ONNX, builds a FP16 TensorRT engine, and reports median latency over 1000 forward passes. Engines are saved under runs/trt_engines/.

---

🎨 Visualisation

visualize.py is a standalone script (no LeAD-M3D imports needed) that renders 3D bounding boxes on KITTI images given prediction files:

python visualize.py \
    --pred   runs/kitti_eval/preds/ \
    --images /path/to/kitti/training/image_2/ \
    --calib  /path/to/kitti/training/calib/ \
    --output vis_output/ \
    --score  0.3

---

📝 Citation

@inproceedings{meier2026leadm3d,
  title     = {{LeAD-M3D}: Leveraging Asymmetric Distillation for Real-Time Monocular {3D} Detection},
  author    = {Meier, Johannes and Michel, Jonathan and Dhaouadi, Oussema and Yang, Yung-Hsu
               and Reich, Christoph and Bauer, Zuria and Roth, Stefan and Pollefeys, Marc
               and Kaiser, Jacques and Cremers, Daniel},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026},
}

---

🙏 Acknowledgements

This codebase builds on YOLOv10 (Wang et al., 2024) and MonoLSS (Li et al., 2024). We thank the authors for their open-source contributions.