topBanner

 

Homepage Evaluation results Download the software, dataset and ground truth Submit your results
Overall results Trial 1 results Trial 2 results Trial 3 results Trial 4 results Trial 5 results Trial 6 results Trial 7 results Trial 8 results

 

Overall evaluation results of trackers

 

Trial-wise performance comparison using Combined Tracking Performance Score (CoTPS)

 

Mean CoTPS of trackers on each trial (P1, P2, ..., P8)
Tracker P1 P2 P3 P4 P5 P6 P7 P8
Boost [2] 0.573 0.576 0.607 0.662 0.535 0.573 0.537 0.541
SemiBoost [3] 0.613 0.627 0.626 0.699 0.587 0.693 0.611 0.627
BeyondSemiBoost [4] 0.601 0.633 0.650 0.806 0.585 0.729 0.613 0.607
MS [1] 0.549 0.537 0.565 0.712 0.571 0.594 0.488 0.515
FragTrack [5] 0.724 0.731 0.725 0.724 0.698 0.748 0.694 0.621
PF [6] 0.593 0.623 0.619 0.762 0.644 0.648 0.590 0.595

 

 

Mean CoTPS Trials

 

Mean CoTPS of trackers on each trial (P1, P2, ..., P8) with all targets; Boost [2] (red), SemiBoost [3] (green), BeyondSemiBoost [4] (blue), MS [1] (black), FragTrack [5] (cyan) and PF [6] (magenta).

 

 

 

Boxplot CoTPS trials

 

CoTPS of trackers on each trial computed with all targets; Boost [2] (red), SemiBoost [3] (green), BeyondSemiBoost [4] (blue), MS [1] (black), FragTrack [5] (cyan) and PF [6] (magenta).

 

 

Target-wise performance comparison using Combined Tracking Performance Score (CoTPS)

 

Mean CoTPS of trackers on each target (H1, H2, ..., H6)
Tracker H1 H2 H3 H4 H5 H6
Boost [2] 0.506 0.529 0.605 0.783 0.563 0.499
SemiBoost [3] 0.415 0.650 0.631 0.915 0.576 0.581
BeyondSemiBoost [4] 0.461 0.642 0.630 0.861 0.706 0.537
MS [1] 0.485 0.592 0.509 0.656 0.566 0.542
FragTrack [5] 0.533 0.660 0.864 0.819 0.927 0.479
PF [6] 0.521 0.519 0.724 0.723 0.575 0.677

 

 

Mean CoTPS targets

 

Mean CoTPS of trackers on each target (H1, H2, ..., H6) with all trials; Boost (red), SemiBoost (green), BeyondSemiBoost (blue), MS (black), FragTrack (cyan) and PF (magenta).

 

 

 

Boxplots CoTPS targets

 

CoTPS of trackers on each target computed with all trials; Boost (red), SemiBoost (green), BeyondSemiBoost (blue), MS (black), FragTrack (cyan) and PF (magenta).

 

 

Cumulative performance of trackers

 

Cumulative performance

 

Cumulative performance of trackers. Mean CoTPS of trackers computed on all trials with all targets are shown with a 'x' in the corresponding boxplots.

 

 

Qualitative tracking results

 

Qtoni Qemilio QvehH3 QvehH4 QperH5 QperH6

(a)

(b)

(c)

(d)

(e)

(f)

 

Visualization of trackers' results in some test frames of test sequences. Boost: red; SemiBoost: green; BeyondSemiBoost: blue; MS: black; FragTrack: cyan; PF: magenta. (a) Visualization of trackers' results in frames (104, 105, 106, 107) involving pose changes of the target (H1). MS, FragTrack, PF cope with the pose change better than Boost, SemiBoost, BeyondSemiBoost. (b) Visualization of trackers' results on H2 in some test frames (frames 145, 280, 487, 511). Frame 145 (first row) shows a 360 degrees turning of target where MS, PF and Boost can track. Frame 280 (second row) shows a scale change. Frames 487 (third row) and 511 (fourth row) involve partial occlusions of the target. PF copes well with both occlusions. BeyondSemiBoost handles well the occlusion in frame 511 only. The remaining trackers have failed in both occlusions. (c) Visualization of trackers' results on H3 in some test frames (6, 21, 82, 159) containing a gradual change in its scale and pose where MS tracks well followed by PF. (d) Visualization of trackers' results on H4 in some test frames (9, 39, 55, 125) containing scale changes. All trackers have struggled to track this target with MS showing the best tracking followed by PF. (e) Visualization of trackers' results on H5 in some test frames (3, 26, 47, 51). Unlike remaining boosting-based trackers, SemiBoost can not distinguish between targets of the same class and tracks a wrong target in frame 26 (second row). PF copes well with a severe occlusion (frames 47, 51). (f) Visualization of trackers' results on H6 in some test frames (49, 311, 359, 616). Unlike remaining boosting-based trackers, SemiBoost, again, can not distinguish targets of the same class and tracks a wrong target in frame 311 (second row). FragTrack has handled the partial occlusion in frame 359 (third row) and slight pose changes in frames 311 and 359 (second and third rows). PF can also track a small part of the target in frame 359.

 

 

References

 

[1] D. Comaniciu, V. Ramesh, and P. Meer, "Kernel-based object tracking", IEEE Trans. on Pattern Analysis and Machine Intelligence, vol. 25, no. 5, pp. 564-577, May 2003.

[2] H. Grabner and H. Bischof, "On-line boosting and vision", in IEEE Conf. on Computer Vision and Pattern Recognition, New York, 2006, pp. 260-267.

[3] H. Grabner, C. Leistner, and H. Bischof, "Semi-supervised on-line boosting for robust tracking", in European Conf. on Computer Vision, Heidelberg, 2008, pp. 234-247.

[4] S. Stalder, H. Grabner, and L. van Gool, "Beyond semi-supervised tracking: Tracking should be as simple as detection, but not simpler than recognition", in Int. Conf. on Computer Vision Workshops, Kyoto, 2009, pp. 1409-1416.

[5] A. Adam, E. Rivlin, and I. Shimshoni, "Robust fragments-based tracking using the integral histogram", in IEEE Conf. on Computer Vision and Pattern Recognition, New York, 2006, pp. 798-805.

[6] P. Prez, C. Hue, J. Vermaak, and M. Gangnet, "Color-based probabilistic tracking", in European Conf. on Computer Vision, Copenhagen, 2002, pp. 661-675.