Home

Awesome

PVO: Panoptic Visual Odometry

Project Page | Paper

<br/>

PVO: Panoptic Visual Odometry

[Weicai Ye, Xinyue Lan]<sup>Co-Authors</sup>, Shuo Chen, Yuhang Ming, Xinyuan Yu, Hujun Bao, Zhaopeng Cui, Guofeng Zhang

CVPR 2023

demo_vid

Test on vkitti 15-deg-left datasets.

  1. prepare. follow prepare.md
conda activate droidenv
  1. generate inital panoptic segmentation.
sh tools/initial_segmentation.sh  
  1. vps->vo,vo Module generate pose, flow and depth.
sh tools/test_vo_scene.sh  
  1. vo->vps, vps Module use flow and depth from vo Module and generate final video panoptic segmentation results and vpq.
sh tools/test_vps.sh  

Metrics on Virtual_KITTI2

SceneRMSEvpq_all/vpq_thing/vpq_stuff
Scene010.37140.39/26.43/44.57
Scene020.05868.84/88.83/62.18
Scene060.11366.38/79.99/62.97
Scene180.95168.35/83.86/63.92
Scene203.50335.11/16.83/40.59

You can get the results in the paper by iterating multiple times.

Train on vkitti 15-deg-left datasets.

  1. To train VPS_Module, you can refer to Detectron2 for more training details. Here for example, you can train vkitti 15-deg-left on 4 GPUs, and training results are saved on output/vps_training/. You can modify the config-file according to the hardware conditions.
python3 -W ignore VPS_Module/tools/train_net.py \
--config-file VPS_Module/configs/COCO-PanopticSegmentation/panoptic_fpn_R_50_3x_vkitti_511.yaml --num-gpu 4 \
MODEL.WEIGHTS checkpoints/panFPN.pth \
OUTPUT_DIR output/vps_training/
  1. To train VO_Module, you can refer to DROID-SLAM for more training details. Here for example, you can train vkitti on 4 GPUs.
python VO_Module/train.py --gpus=4 --lr=0.00025

Visualization

You can refer to DROID-SLAM for visualization. All demos can be run on a GPU with 11G of memory. While running, press the "s" key to increase the filtering threshold (= more points) and "a" to decrease the filtering threshold (= fewer points).

python VO_Module/evaluation_scripts/test_vo.py --datapath=datasets/Virtual_KITTI2/Scene01 --segm_filter True 

Citation

If you find this code useful for your research, please use the following BibTeX entry.


@inproceedings{Ye2023PVO,
  title={{PVO: Panoptic visual odometry}},
  author={Ye, Weicai and Lan, Xinyue and Chen, Shuo and Ming, Yuhang and Yu, Xingyuan and Bao, Hujun and Cui, Zhaopeng and Zhang, Guofeng},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={9579--9589},
  year={2023}
}

Acknowledgement

Some code snippets are borrowed from DROID-SLAM and Detectron2. Thanks for these great projects.