Released code for Objects are Different: Flexible Monocular 3D Object Detection, CVPR21.

Work in progress.


This repo is tested with Ubuntu 20.04, python3.7, pytorch1.4.0 and cuda==10.1

conda create -n monoflex python=3.7

conda activate monoflex

Install PyTorch and other dependencies:

conda install pytorch==1.4.0 torchvision==0.5.0 cudatoolkit=10.1 -c pytorch

pip install -r requirements.txt

Build DCNv2 and the project

cd models/backbone/DCNv2


cd ../../..

python setup develop

Data Preparation

Please download KITTI dataset and organize the data as follows:


Then modify the paths in config/ according to your data path.

Training & Evaluation

Training with one GPU. (TODO: The multi-GPU training will be further tested.)

CUDA_VISIBLE_DEVICES=0 python tools/ --batch_size 8 --config runs/monoflex.yaml --output output/exp

The model will be evaluated periodically (can be adjusted in the CONFIG) during training and you can also evaluate a checkpoint with

CUDA_VISIBLE_DEVICES=0 python tools/ --config runs/monoflex.yaml --ckpt YOUR_CKPT  --eval

You can also specify --vis when evaluation to visualize the predicted heatmap and 3D bounding boxes. The pretrained model for train/val split and logs are here.

Note: we observe an obvious variation of the performance for different runs and we are still investigating possible solutions to stablize the results, though it may inevitably due to the utilized uncertainties.


If you find our work useful in your research, please consider citing:

    author    = {Zhang, Yunpeng and Lu, Jiwen and Zhou, Jie},
    title     = {Objects Are Different: Flexible Monocular 3D Object Detection},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2021},
    pages     = {3289-3298}


The code is heavily borrowed from SMOKE and thanks for their contribution.