MonoDGC: Dynamic Graph Cross-Former for Monocular 3D Object Detection

This repository hosts the official implementation of MonoDGC: Dynamic Graph Cross-Former for Monocular 3D Object Detection based on the excellent work MonoDGP.

Results

You can find the Val results in MonoDGC/monodgc_logs/ And the Test results submitted to the official KITTI Benchmark: https://www.cvlibs.net/datasets/kitti/eval_object_detail.php?&result=8380e006c8842386e51232717ccc6790be5e7f78

Dataset Structure

Download KITTI datasets on https://www.cvlibs.net/datasets/kitti/eval_object.php?obj_benchmark=3d and prepare the directory structure as:

│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/monodgc.yaml. The "tools/create_data.py" script can be used to configure the format of the KITTI dataset.

python  tools/create_data.py kitti \
            --root-path ./data/kitti \
            --out-dir ./data/kitti \
            --extra-tag kitti

Install Dependencies

git clone https://github.com/Tracygc/MonoDGC.git
conda create --name mmdet3d python=3.9
conda activate mmdet3d 
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128  -i https://pypi.tuna.tsinghua.edu.cn/simple

For configuring a Linux environment with the RTX 5090 GPU and CUDA 12.8, you can refer to my blog post at this link: https://blog.csdn.net/TracyGC/article/details/156197710?spm=1001.2014.3001.5501

Get Started

Train

You can train MonoDGC here, and you can modify the settings of models and training in configs/monodgc.yaml:

CUDA_VISIBLE_DEVICES=0 nohup python tools/train_val7.py --config configs/monodgc.yaml > logs/monodgc.log 2>&1

You can also run MonoDGP here:

CUDA_VISIBLE_DEVICES=0 nohup python tools/train_val.py --config configs/monodgp.yaml > logs/monodgp.log 2>&1

Test

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

CUDA_VISIBLE_DEVICES=0 nohup python tools/train_val7.py --config configs/monodgc.yaml -e > logs/monodgc_test.log 2>&1

Test the inference time on your own device:

python tools/test_runtime.py

Node classification and Graph classification

You can refer to MonoDGC/Cross-Former in detail.

Reference Links

https://github.com/PuFanqi23/MonoDGP
https://github.com/ZrrSkywalker/MonoDETR
https://github.com/SuperMHP/GUPNet_Plus