CVPR 2019 Tracking Challenge Results

Click on a measure to sort the table accordingly. See below for a more detailed description.


Showing only entries that use public detections!


Benchmark Statistics

TrackerMOTAIDF1 MOTPMTMLFPFNRecallPrecisionFAFID Sw.FragHz
Tracktor++
1. online method using public detections
51.3
±18.7
47.6
±12.1
76.7 313 (24.9)326 (26.0)16,263 253,680 54.7 95.0 3.6 2,584 (47.2)4,824 (88.2)2.7
P. Bergmann, T. Meinhardt, L. Leal-Taixé. Tracking without bells and whistles. In ICCV, 2019.
HAM_HI
2. online method using public detections
43.0
±24.0
43.6
±12.3
77.1 353 (28.1)274 (21.8)72,018 243,055 56.6 81.5 16.1 4,153 (73.4)4,801 (84.8)0.8
Y. Yoon, A. Boragule, Y. Song, K. Yoon, M. Jeon. Online Multi-Object Tracking with Historical Appearance Matching and Scene Adaptive Detection Filtering. In IEEE AVSS, 2018.
Seq2Seq
3. using public detections
39.7
±16.7
37.6
±10.0
77.2 179 (14.3)343 (27.3)38,991 292,239 47.8 87.3 8.7 6,408 (134.0)10,254 (214.4)0.4
Anonymous submission
DD_TAMA19
4. online method using public detections
47.6
±20.3
48.7
±13.2
77.6 342 (27.2)297 (23.6)38,194 252,934 54.8 88.9 8.5 2,437 (44.4)3,887 (70.9)0.2
Y. Yoon, D. Kim, K. Yoon, Y. Song, M. Jeon. Online Multiple Pedestrian Tracking using Deep Temporal Appearance Matching Association. In arXiv:1907.00831, 2019.
V_IOU
5. using public detections
46.7
±19.6
46.0
±12.4
78.2 288 (22.9)306 (24.4)33,776 261,964 53.2 89.8 7.5 2,589 (48.6)4,354 (81.8)18.2
E. Bochinski, T. Senst, T. Sikora. Extending IOU Based Multi-Object Tracking by Visual Information. In IEEE International Conference on Advanced Video and Signals-based Surveillance, 2018.
IOU_19
6. using public detections
35.8
±15.1
25.7
±8.6
78.9 126 (10.0)389 (31.0)24,427 319,696 42.9 90.8 5.5 15,676 (365.3)17,864 (416.3)183.3
E. Bochinski, V. Eiselein, T. Sikora. High-Speed Tracking-by-Detection Without Using Image Information. In International Workshop on Traffic and Street Surveillance for Safety and Security at IEEE AVSS 2017, 2017.
SRK_ODESA
7. online method using public detections
54.8
±20.3
52.2
±13.3
79.0 444 (35.4)241 (19.2)33,814 215,572 61.5 91.1 7.5 3,750 (61.0)5,493 (89.3)1.2
D. Borysenko, D. Mykheievskyi, V. Porokhonskyy. ODESA: Object Descriptor that is Smooth Appearance-wise for object tracking tasks. In (to be submitted to ECCV'20), .
SequencesFramesTrajectoriesBoxes
444791492803370

Difficulty Analysis

Sequence difficulty (from easiest to hardest, measured by average MOTA)

CVPR19-04

CVPR19-04

(61.2 MOTA)

CVPR19-07

CVPR19-07

(51.2 MOTA)

CVPR19-06

CVPR19-06

(24.7 MOTA)

CVPR19-08

CVPR19-08

(14.1 MOTA)


Evaluation Measures

Lower is better. Higher is better.
Measure Better Perfect Description
MOTA higher 100 % Multiple Object Tracking Accuracy [1]. This measure combines three error sources: false positives, missed targets and identity switches.
MOTP higher 100 % Multiple Object Tracking Precision [1]. The misalignment between the annotated and the predicted bounding boxes.
IDF1 higher 100 % ID F1 Score [2]. The ratio of correctly identified detections over the average number of ground-truth and computed detections.
FAF lower 0 The average number of false alarms per frame.
MT higher 100 % Mostly tracked targets. The ratio of ground-truth trajectories that are covered by a track hypothesis for at least 80% of their respective life span.
ML lower 0 % Mostly lost targets. The ratio of ground-truth trajectories that are covered by a track hypothesis for at most 20% of their respective life span.
FP lower 0 The total number of false positives.
FN lower 0 The total number of false negatives (missed targets).
ID Sw. lower 0 The total number of identity switches. Please note that we follow the stricter definition of identity switches as described in [3].
Frag lower 0 The total number of times a trajectory is fragmented (i.e. interrupted during tracking).
Hz higher Inf. Processing speed (in frames per second excluding the detector) on the benchmark. The frequency is provided by the authors and not officially evaluated by the MOTChallenge.

Legend

Symbol Description
online method This is an online (causal) method, i.e. the solution is immediately available with each incoming frame and cannot be changed at any later time.
using public detections This method used the provided detection set as input.
using public detections This method used the provided detection set as input.
new This entry has been submitted or updated less than a week ago.
new This entry has been submitted or updated less than a week ago.

References:


[1] Bernardin, K. & Stiefelhagen, R. Evaluating Multiple Object Tracking Performance: The CLEAR MOT Metrics. Image and Video Processing, 2008(1):1-10, 2008.
[2] Ristani, E., Solera, F., Zou, R., Cucchiara, R. & Tomasi, C. Performance Measures and a Data Set for Multi-Target, Multi-Camera Tracking. In ECCV workshop on Benchmarking Multi-Target Tracking, 2016.
[3] Li, Y., Huang, C. & Nevatia, R. Learning to associate: HybridBoosted multi-target tracker for crowded scene. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2009.