This repository provides the official implementation of the TrackFormer: Multi-Object Tracking with Transformers paper by Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe and Christoph Feichtenhofer. The codebase builds upon DETR, Deformable DETR and Tracktor.
The challenging task of multi-object tracking (MOT) requires simultaneous reasoning about track initialization, identity, and spatiotemporal trajectories. We formulate this task as a frame-to-frame set prediction problem and introduce TrackFormer, an end-to-end MOT approach based on an encoder-decoder Transformer architecture. Our model achieves data association between frames via attention by evolving a set of track predictions through a video sequence. The Transformer decoder initializes new tracks from static object queries and autoregressively follows existing tracks in space and time with the new concept of identity preserving track queries. Both decoder query types benefit from self- and encoder-decoder attention on global frame-level features, thereby omitting any additional graph optimization and matching or modeling of motion and appearance. TrackFormer represents a new tracking-by-attention paradigm and yields state-of-the-art performance on the task of multi-object tracking (MOT17) and segmentation (MOTS20).
We refer to our docs/INSTALL.md for detailed installation instructions.
We refer to our docs/TRAIN.md for detailed training instructions.
In order to evaluate TrackFormer on a multi-object tracking dataset, we provide the src/track.py script which supports several datasets and splits interchangle via the dataset_name argument (See src/datasets/tracking/factory.py for an overview of all datasets.) The default tracking configuration is specified in cfgs/track.yaml. To facilitate the reproducibility of our results, we provide evaluation metrics for both the train and test set.
python src/track.py with reid
python src/track.py with \
reid \
public_detections=min_iou_0_5 \
obj_detect_checkpoint_file=models/mot17_deformable_multi_frame/checkpoint_epoch_50.pth
python src/track.py with \
reid \
obj_detect_checkpoint_file=models/mot20_crowdhuman_deformable_multi_frame/checkpoint_epoch_50.pth
python src/track.py with \
dataset_name=MOTS20-ALL \
obj_detect_checkpoint_file=models/mots20_train_masks/checkpoint.pth
Our tracking script only applies MOT17 metrics evaluation but outputs MOTS20 mask prediction files. To evaluate these download the official MOTChallengeEvalKit.
| MOTS20 | sMOTSA | IDF1 | FP | FN | IDs | | :---: | :---: | :---: | :---: | :---: | :---: | | **Train** | -- | -- | -- | -- | -- | | **Test** | 54.9 | 63.6 | 2233 | 7195 | 278 |To facilitate the application of TrackFormer, we provide a demo interface which allows for a quick processing of a given video sequence.
ffmpeg -i data/snakeboard/snakeboard.mp4 -vf fps=30 data/snakeboard/%06d.png
python src/track.py with \
dataset_name=DEMO \
data_root_dir=data/snakeboard \
output_dir=data/snakeboard \
write_images=pretty
If you use this software in your research, please cite our publication:
@InProceedings{meinhardt2021trackformer,
title={TrackFormer: Multi-Object Tracking with Transformers},
author={Tim Meinhardt and Alexander Kirillov and Laura Leal-Taixe and Christoph Feichtenhofer},
year={2021},
eprint={2101.02702},
archivePrefix={arXiv},
}