Skip to content

Kemo-Huang/BiTrack

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

6 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

BiTrack

Bidirectional Offline 3D Multi-Object Tracking using Camera-LiDAR Data

Get Started

Put KITTI tracking data to the "data/kitti/tracking" directory (symbolic links are recommended).

  • If you want to generate 2D-3D detection results using VoxelRCNN and SpatialEmbedding, follow the KITTI detection instructions.

  • If you have generated detection files using other models, skip to 2D-3D fusion instructions. Put 3D detection results to the "data/kitti/tracking/$split/det3d_out/$det3d_name" directory and put 2D segmentation results to the "data/kitti/tracking/$split/seg_out/$seg_name" directory. You can also use other file paths but the configuration file should be changed accordingly.

KITTI Detection

KITTI 3D Object Detection

  1. Build CUDA operators for 3D-IoU and PointNet++.
python setup.py develop
  1. Download the converted model weight file from Google Drive and put it to "detection/voxel_rcnn/voxel_rcnn.pth".

  2. Network inference using a specific configuration file under the "configs" directory.

python kitti_3d_detection.py $config_path $split --inference
  1. (Optional) Average precision evaluation for cars. (1) Convert the tracking labels to the detection format. (2) Convert the detection results to the detection format. (3) Perform evaluation using the converted labels and results.
python data_processing/dataset_tracking2object.py
python data_processing/tracking2object.py $result_src $result_dst
python eval_kitti_detection.py ./data/kitti/detection/training/label_2 $result_dst

KITTI 2D Instance Segmentation

  1. Download the converted model weight file from Google Drive and put it to "segmentation/spatial_embeddings/spatial_embeddings.pth".

  2. Network inference using a specific configuration file under the "configs" directory.

python kitti_2d_mots.py $config_path $split

KITTI 2D-3D Fusion

  1. Crop LiDAR points that are inside 3D bounding boxes.
python data_processing/crop_points.py $config_path $split
  1. Save image shapes to json (not all KITTI images have the same shape).
python data_processing/save_img_shapes.py $config_path $split
  1. Run the detection fusion script.
python kitti_2d_3d_det_fusion.py $config_path $split

KITTI Tracking

  1. Forward tracking.
python kitti_3d_tracking.py $config_path $forward_tag $split
  1. Backward tracking.
python kitti_3d_tracking.py $config_path $backward_tag $split --backward
  1. Trajectory fusion and refinement.
python kitti_trajectory_refinement.py $config_path $final_tag $split $foward_tag $backward_tag
  1. Evaluation for cars.
python TrackEval/scripts/run_kitti.py --TIME_PROGRESS False --PRINT_CONFIG False --GT_FOLDER data/kitti/tracking/training --TRACKERS_FOLDER output/kitti/$split --CLASSES_TO_EVAL car --TRACKERS_TO_EVAL $tag --SPLIT_TO_EVAL $split

About

Official PyTorch implementation of BiTrack

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published