This project is used to implement the KITTI object detection and tracking system using a pretrained CenterNet model.
Firstly, download the KITTI left images and labels to evaluate the model.
Secondly, download the pretrained ddd_3dop.pth model.
Finally, follow these steps for installation.
If you have some issues with torch - torchvision compatibility, try this:
sudo pip3 install -r requirements.txt
To predict and track objects from KITTI, use:
python3 predict.py [--dataset_type] [--model_name] [--score_threshold] [--dist_threshold] [--iou_threshold] [--depth_threshold] [--check_zmin] [--check_dim_ratio] [--ttl] [--begin_index] [--end_index] [--show_frames/no_show_frames] [--verbose/no_verbose] [--with_keys/with_no_keys]
To show already predicted objects, use:
python3 parse.py [--dataset_type] [--index] [--is_gt]
Screenshots were taken from the first image of 0001 training dataset. Raw 2D bounding boxes:
Tracked 3D bounding boxes:
Bird-view bounding boxes:
Unfortunately, this model can't be evaluated on the testing KITTI dataset due to its policy. So, only training dataset's results are published.
| Class | MOTA | MOTP | MT | ML | IDS | FRAG |
|---|---|---|---|---|---|---|
| Car | 79.90% | 80.22% | 70.92% | 6.91% | 165 | 539 |
| Pedestrian | 52.26% | 69.23% | 39.52% | 11.97% | 490 | 915 |