Code for the TrackR-CNN baseline for the Multi Object Tracking and Segmentation (MOTS) task.
Javid Haji-zada, Anar Huseynov, and Leyla Ismayilova
https://www.vision.rwth-aachen.de/media/papers/mots-multi-object-tracking-and-segmentation/MOTS.pdf
https://github.com/VisualComputingInstitute/mots_tools
You'll need to install the following packages (possibly more):
tensorflow-gpu pycocotools numpy scipy sklearn pypng opencv-python munkres
In particular, the code has been tested with Python 3.6.7 and Tensorflow-GPU 1.14.0 running on a single GTX 1080 Ti gpu.
Furthermore, you'll need the KITTI MOTS dataset, where we assume you have a folder /path/to/kitti_mots with subfolders /path/to/kitti_mots/images containing the input images (i.e. there exist subfolders /path/to/kitti_mots/images/0000, /path/to/kitti_mots/images/0001, ...) and /path/to/kitti_mots/instances (again with subfolders 0000, 0001, ...) containing the annotations.
Also, create the following directories for logs, model files etc. in the base directory of the repository:
mkdir forwarded models summaries logs
Pre-trained models can be downloaded here: https://omnomnom.vision.rwth-aachen.de/data/trackrcnn/
In the configuration files, you'll need to adjust the KITTI_segtrack_data_dir and load_init flags to point to the KITTI MOTS data directory and the path to the pretrained model, respectively. Logs, checkpoints and summaries are stored in the logs/, models/ and summaries/ subdirectories.
So all in all, your folder structure should look like this:
data/
- KITTI_MOTS/
-- train/
--- images/
---- 0000/
----- 000000.png
----- 000001.png
----- ...
---- 0001/
---- ...
--- instances/
---- 0000/
----- 000000.png
----- 000001.png
----- ...
---- 0001/
---- ...
models/
- conv3d_sep2/
-- conv3d_sep2-00000005.data-00000-of-00001
-- conv3d_sep2-00000005.index
-- conv3d_sep2-00000005.meta
- converted.data-00000-of-00001
- converted.meta
- converted.index
...
main.py
So point KITTI_segtrack_data_dir to data/KITTI_MOTS/train/ and load_init to models/converted.
In order to train a model, run main.py with the corresponding configuration file. For the baseline model with two separable 3D convolutions and data association with learned embeddings, use
python main.py configs/conv3d_sep2
Either first train your own model as described above, or download previously trained and extract the files into models/conv3d_sep2/
To obtain the model's predictions (we call this "forwarding") run:
python main.py configs/conv3d_sep2 "{\"task\":\"forward_tracking\",\"dataset\":\"KITTI_segtrack_feed\",\"load_epoch_no\":5,\"batch_size\":5,\"export_detections\":true,\"do_tracking\":false,\"video_tags_to_load\":[\"0002\",\"0006\",\"0007\",\"0008\",\"0010\",\"0013\",\"0014\",\"0016\",\"0018\",\"0000\",\"0001\",\"0003\",\"0004\",\"0005\",\"0009\",\"0011\",\"0012\",\"0015\",\"0017\",\"0019\",\"0020\"]}"
The json string supplied as an additional argument here overwrites the settings in the config file. Use video_tags_to_load to obtain predictions for specific sequences (in the example, all KITTI MOTS sequences are chosen). Output is written to the forwarded/ subdirectory.
The model predictions as obtained by the previous command are not yet linked over time. You can use the following command to run the tracking algorithm described in the paper and to obtain final results in the forwarded/ subdirectory which can be processed by the mots_tools scripts:
python main.py configs/conv3d_sep2
"{\"build_networks\":false,\"import_detections\":true,\"task\":\"forward_tracking\",\"dataset\":\"KITTI_segtrack_feed\",\"do_tracking\":true,\"visualize_detections\":false,\"visualize_tracks\":false,\"load_epoch_no\":5,\"video_tags_to_load\":[\"0002\",\"0006\",\"0007\",\"0008\",\"0010\",\"0013\",\"0014\",\"0016\",\"0018\"]}"
You can also visualize the tracking results here by setting visualize_tracks to true.
The script for random tuning will find the best combination of tracking parameters on the training set and then evaluate these parameters on the validation set. This is how the results in the MOTS paper are obtained.
To use this script, run
python scripts/eval/segtrack_tune_experiment.py /path/to/detections/ /path/to/groundtruth/ /path/to/precomputed_optical_flow/ /path/to/output_file /path/to/tmp_folder/ /path/to/mots_eval/ association_type num_iterations
where /path/to/detections/ is a folder containing the model output on the training set (obtained by the forwarding command above); /path/to/mots_eval/ refers to the official evaluation script (link see above, clone the repository and supply the path here); association_type determines the method for associating detections into tracks and is either reid (using the association embeddings - use this if unsure!), mask (using mask warping), bbox_iou (using bounding box warping with median optical flow) or bbox_center (nearest neighbor matching); num_iterations is the number of random trials (1000 in the paper); /path/to/groundtruth/ refers to the instances or instances_txt folder containing the annotations (which you can download from the project website); /path/to/precomputed_optical_flow can usually be set to a dummy folder (it refers to optical flow for input images, which is used only for a few of the ablation experiments in the paper - namely when setting association_type to mask or bbox_iou but if you set it to something else, then the flow path is ignored); at /path/to/output_file, a file will be created containing the results of the individual tuning iterations, please make sure this path is writable; at /path/to/tmp_folder a lot of intermediate folders will be stored, you can delete these afterwards.
Parts of this code are based on Tensorpack (https://github.com/tensorpack/tensorpack/tree/master/examples/FasterRCNN) and RETURNN (https://github.com/rwth-i6/returnn/blob/master/Log.py).
Citation
@inproceedings{Voigtlaender19CVPR_MOTS,
author = {Paul Voigtlaender and Michael Krause and Aljosa Osep and Jonathon Luiten and Berin Balachandar Gnana Sekar and Andreas Geiger and Bastian Leibe},
title = {{MOTS}: Multi-Object Tracking and Segmentation},
booktitle = {CVPR},
year = {2019},
}