Pseudo-Stereo for Monocular 3D Object Detection in Autonomous Driving. (CVPR 2022)

Authors: Yi-Nan Chen, Hang Dai and Yong Ding
[Paper] [Supplementary file]

The code is tested on an ubuntu server with NVIDIA RTX 3090.

We now release the code for feature-level generation and faeture-clone generation. We apply our methods on the follows stereo-based detectors:

LIGA-Stereo

Step I: Follow the instruction of LIGA-Stereo to install the dependencies.

Step II: Replace the some files in LIGA-Stereo use the files that we provide in here.

Step III: Prepare the data. Please fisrt follow instruction in LIGA-Stereo to prepeare the data. Then download the estimated depth maps by DORN from here(training, testing). Then put the depth maps into data/training/depth_2_dorn.

Step IV: Training, use the command as follows to train the model. Note that we can only set bacth size to 1 on each GPU in our practice.
- feature-level generation
./scripts/dist_train.sh ${NUM_GPUS} 'exp_name' ./configs/stereo/kitti_models/feature_level_generation.yaml
- feature-clone
./scripts/dist_train.sh ${NUM_GPUS} 'exp_name' ./configs/stereo/kitti_models/feature_clone.yaml
YOLOStereo3D

Step I: Follow the instruction of visualDet3D to install the dependencies.

Step II: Replace the some files in visualDet3D use the files that we provide in here.

Step III: Prepare the data. Please fisrt follow instruction in visualDet3D to prepeare the data. Then download the estimated depth maps by DORN from here(training, testing). Then put the depth maps into data/training/image_2_dorn.

Step IV: Training, use the command as follows to train the model.
- feature-level generation
./launcher/train.sh --config/feature_level_generation.py 0 $experiment_name

For image-level generation, we will release the synthesised virtual right iamges.

Citation

@InProceedings{Chen_2022_CVPR,
    title={Pseudo-Stereo for Monocular 3D Object Detection in Autonomous Driving},
    author={Yi-Nan Chen and Hang Dai and Yong Ding},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2022},
}