DDMP-3D
Pytorch implementation of Depth-conditioned Dynamic Message Propagation forMonocular 3D Object Detection, a paper on CVPR2021.
Instroduction
The objective of this paper is to learn context- and depthaware feature representation to solve the problem of monocular 3D object detection. We make following contributions: (i) rather than appealing to the complicated pseudo-LiDAR based approach, we propose a depth-conditioned dynamic message propagation (DDMP) network to effectively integrate the multi-scale depth information with the image context; (ii) this is achieved by first adaptively sampling context-aware nodes in the image context and then dynamically predicting hybrid depth-dependent filter weights and affinity matrices for propagating information; (iii) by augmenting a center-aware depth encoding (CDE) task, our method successfully alleviates the inaccurate depth prior; (iv) we thoroughly demonstrate the effectiveness of our proposed approach and show state-of-the-art results among the monocular-based approaches on the KITTI benchmark dataset.
Requirements
Installation
Our code is based on DGMN, please refer to the installation for maskrcnn-benchmark compilation.
-
My settings
conda activate maskrcnn_benchmark (maskrcnn_benchmark) conda list python 3.8.5 pytorch 1.4.0 cudatoolkit 10.0.130 torchfile 0.1.0 torchvision 0.5.0 apex 0.1
Data preparation
Download and unzip the full KITTI detection dataset to the folder /path/to/kitti/. Then place a softlink (or the actual data) in data/kitti/. There are two widely used training/validation set splits for the KITTI dataset. Here we only show the setting of split1, you can set split2 accordingly.
cd D4LCN
ln -s /path/to/kitti data/kitti
ln -s /path/to/kitti/testing data/kitti_split1/testing
Our method uses DORN (or other monocular depth models) to extract depth maps for all images. You can download and unzip the depth maps extracted by DORN here and put them (or softlink) to the folder data/kitti/depth_2/. (You can also change the path in the scripts setup_depth.py). Additionally, we also generate the xyz map (xy are the values along x and y axises on 2D plane, and z is the depth value) and save as pickle files and then operate like depth map.
Then use the following scripts to extract the data splits, which use softlinks to the above directory for efficient storage.
python data/kitti_split1/setup_split.py
python data/kitti_split1/setup_depth.py
Next, build the KITTI devkit eval for split1.
sh data/kitti_split1/devkit/cpp/build.sh
Lastly, build the nms modules
cd lib/nms
make
Training
You can change the batch_size according to the number of GPUs, default: 8 GPUs with batch_size = 5 on Tesla v100(32G).
If you want to utilize the resnet backbone pre-trained on the COCO dataset, it can be downloaded from git or Google Drive, default: ImageNet pretrained pytorch model, we downloaded the model and saved at 'data/'. You can also set use_corner and corner_in_3d to False for quick training.
See the configurations in scripts/config/config.py and scripts/train.py for details.
sh train.sh
Testing
Generate the results using:
python scripts/test.py
we afford the generated results for evaluation due to the tedious process of data preparation process. Unzip the output.zip and then execute the above evaluation commonds. We show the results in paper, and supplementary. Additionally, we also trained a model replacing the depth map (only contains value of z) with coordinate xyz (xy are the values along x and y axises on 2D plane), which achieves the best performance. You can download the best model on Google Drive.
Models | AP3D11@mod. | AP3D11@easy | AP3D11@hard |
---|---|---|---|
model in paper | 23.13 / 27.46 | 31.14 / 37.71 | 19.45 / 24.53 |
model in supp | 23.17 / 27.85 | 32.40 / 42.05 | 19.35 / 24.91 |
model with coordinate(xyz), config | 23.53 / 28.16 | 30.21 / 38.78 | 19.72 / 24.80 |
Acknowledgements
We thank D4LCN and DGMN for their great works and repos.
Citation
If you find this project useful in your research, please consider citing:
@inproceedings{wang2021depth,
title={Depth-conditioned Dynamic Message Propagation for Monocular 3D Object Detection},
author={Wang, Li and Du, Liang and Ye, Xiaoqing and Fu, Yanwei and Guo, Guodong and Xue, Xiangyang and Feng, Jianfeng and Zhang, Li},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={454--463},
year={2021}
}
Contact
For questions regarding DDMP-3D, feel free to post here or directly contact the authors ([email protected]).