Skip to content

Latest commit

 

History

History

tracktor

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Tracking without Bells and Whistles

Abstract

The problem of tracking multiple objects in a video sequence poses several challenging tasks. For tracking-by-detection, these include object re-identification, motion prediction and dealing with occlusions. We present a tracker (without bells and whistles) that accomplishes tracking without specifically targeting any of these tasks, in particular, we perform no training or optimization on tracking data. To this end, we exploit the bounding box regression of an object detector to predict the position of an object in the next frame, thereby converting a detector into a Tracktor. We demonstrate the potential of Tracktor and provide a new state-of-the-art on three multi-object tracking benchmarks by extending it with a straightforward re-identification and camera motion compensation. We then perform an analysis on the performance and failure cases of several state-of-the-art tracking methods in comparison to our Tracktor. Surprisingly, none of the dedicated tracking methods are considerably better in dealing with complex tracking scenarios, namely, small and occluded objects or missing detections. However, our approach tackles most of the easy tracking scenarios. Therefore, we motivate our approach as a new tracking paradigm and point out promising future research directions. Overall, Tracktor yields superior tracking performance than any current tracking method and our analysis exposes remaining and unsolved tracking challenges to inspire future research directions.

Citation

@inproceedings{bergmann2019tracking,
  title={Tracking without bells and whistles},
  author={Bergmann, Philipp and Meinhardt, Tim and Leal-Taixe, Laura},
  booktitle={Proceedings of the IEEE international conference on computer vision},
  pages={941--951},
  year={2019}
}

We implement Tracktor with independent detector and ReID models. To train a model by yourself, you need to train a detector following here and also train a ReID model following here. The configs in this folder are basically for inference.

Results and models on MOT15

Method Detector ReID Train Set Test Set Public Inf time (fps) MOTA IDF1 FP FN IDSw. Config Download
Tracktor R50-FasterRCNN-FPN R50 half-train half-val Y - 61.8 64.9 1235 6877 116 config detector | detector_log | reid | reid_log
Tracktor R50-FasterRCNN-FPN R50 half-train half-val N - 66.8 68.4 3049 3922 179 config detector | detector_log | reid | reid_log

Results and models on MOT16

Method Detector ReID Train Set Test Set Public Inf time (fps) MOTA IDF1 FP FN IDSw. Config Download
Tracktor R50-FasterRCNN-FPN R50 half-train half-val Y - 54.1 61.5 425 23894 182 config detector | detector_log | reid | reid_log
Tracktor R50-FasterRCNN-FPN R50 half-train half-val N - 63.4 66.2 4175 14911 444 config detector | detector_log | reid | reid_log

Results and models on MOT17

The implementations of Tracktor follow the official practices. In the table below, the result marked with * (the last line) is the official one. Our implementation outperform it by 4.9 points on MOTA and 3.3 points on IDF1.

Method Detector ReID Train Set Test Set Public Inf time (fps) MOTA IDF1 FP FN IDSw. Config Download
Tracktor R50-FasterRCNN-FPN R50 half-train half-val Y 3.2 57.3 63.4 1254 67091 614 config detector reid
Tracktor R50-FasterRCNN-FPN R50 half-train half-val N 3.1 64.1 66.9 11088 45762 1233 config detector reid
Tracktor R50-FasterRCNN-FPN R50 train test Y 3.2 61.2 58.4 8609 207627 2634 config detector reid
Tracktor* R50-FasterRCNN-FPN R50 train test Y - 56.3 55.1 8866 235449 1987 - -
Tracktor
(FP16)
R50-FasterRCNN-FPN R50 half-train half-val N - 64.7 66.6 10710 45270 1152 config detector | detector_log | reid | reid_log

Note:

  • FP16 means Mixed Precision (FP16) is adopted in training.

Results and models on MOT20

The implementations of Tracktor follow the official practices. In the table below, the result marked with * (the last line) is the official one. Our implementation outperform it by 5.3 points on MOTA and 2.1 points on IDF1.

Method Detector ReID Train Set Test Set Public Inf time (fps) MOTA IDF1 FP FN IDSw. Config Download
Tracktor R50-FasterRCNN-FPN R50 half-train half-val Y - 70.6 65.4 3652 175955 1441 config detector | detector_log | reid | reid_log
Tracktor R50-FasterRCNN-FPN R50 half-train half-val N - 70.9 64.1 5539 171653 1619 config detector | detector_log | reid | reid_log
Tracktor R50-FasterRCNN-FPN R50 train test Y - 57.9 54.8 16203 199485 2299 config detector | detector_log | reid | reid_log
Tracktor R50-FasterRCNN-FPN* R50 train test Y - 52.6 52.7 6930 236680 1648 - -

Note: When running demo_mot.py, we suggest you use the config containing private, since private means the MOT method doesn't need external detections.