MonoDGP: Monocular 3D Object Detection with Decoupled-Query and Geometry-Error Priors

This repository hosts the official implementation of MonoDGP: Monocular 3D Object Detection with Decoupled-Query and Geometry-Error Priors based on the excellent work MonoDETR. In this work, we propose a novel transformer-based monocular method called MonoDGP, which adopts geometry errors to correct the projection formula. We also introduce a 2D visual decoder for query initialization and a region segmentation head for feature enhancement.

The official results in the paper:

Models	Val, AP3D|R40
	Easy	Mod.	Hard
MonoDGP	30.7624%	22.3421%	19.0144%

New and better results in this repo:

Models	Val, AP3D|R40			Logs	Ckpts
	Easy	Mod.	Hard
MonoDGP	30.9663%	22.4953%	19.9641%	log	ckpt
	30.7194%	22.6883%	19.4441%	log	ckpt
	30.1314%	22.7109%	19.3978%	log	ckpt

We also provide a new ckpt pretrained on the trainval set, you can directly utilize this ckpt and submit test results to the KITTI 3D object detection benchmark (You need to register an account) to verify our method.

Models	Test, AP3D|R40			Ckpts	Results
	Easy	Mod.	Hard
MonoDGP	26.12%	18.87%	16.79%	ckpt	data

Test results submitted to the official KITTI Benchmark:

Car category:

All categories:

Installation

1. Clone this project and create a conda environment:

   git clone https://github.com/PuFanqi23/MonoDGP.git
   cd MonoDGP
   
   conda create -n monodgp python=3.8
   conda activate monodgp

2. Install pytorch and torchvision matching your CUDA version:

   # For example, We adopt torch 1.9.0+cu111
   pip install torch==1.9.0+cu111 torchvision==0.10.0+cu111 torchaudio==0.9.0 -f https://download.pytorch.org/whl/torch_stable.html

3. Install requirements and compile the deformable attention:

   pip install -r requirements.txt
   
   cd lib/models/monodgp/ops/
   bash make.sh
   
   cd ../../../..

4. Download KITTI datasets and prepare the directory structure as:

   │MonoDGP/
   ├──...
   │data/kitti/
   ├──ImageSets/
   ├──training/
   │   ├──image_2
   │   ├──label_2
   │   ├──calib
   ├──testing/
   │   ├──image_2
   │   ├──calib

   You can also change the data path at "dataset/root_dir" in configs/monodgp.yaml.

Get Started

Train

You can modify the settings of models and training in configs/monodgp.yaml and indicate the GPU in train.sh:

bash train.sh configs/monodgp.yaml > logs/monodgp.log

Test

The best checkpoint will be evaluated as default. You can change it at "tester/checkpoint" in configs/monodgp.yaml:

bash test.sh configs/monodgp.yaml

You can test the inference time on your own device:

python tools/test_runtime.py

Citation

If you find our work useful in your research, please consider giving us a star and citing:

@inproceedings{pu2025monodgp,
  title={Monodgp: Monocular 3D object detection with decoupled-query and geometry-error priors},
  author={Pu, Fanqi and Wang, Yifan and Deng, Jiru and Yang, Wenming},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={6520--6530},
  year={2025}
}

Acknowlegment

This repo benefits from the excellent work MonoDETR.