Online Object Tracking: A Benchmark 论文笔记

Factors that affect the performance of a tracing algorithm

1 Illumination variation
2 Occlusion
3 Background clutters

Main modules for object tracking

1 Target representation scheme
2 Search mechanism
3 Model update

Evaluation Methodology

1 Precison plot:
The percentage of frames whose estimated location is within the given threshold distance of the ground truth.
x coordinate: threshold

2 Success plot: 
The ratios of successful frames at the thresholds varied from 0 to 1
x coordinate: threshold

3 Robustness Evaluation
A OPE: one-pass evaluation
B TRE temporal robustness evaluation
C SRE spatial robustness evaluation

Overall Performance

1  TLD performs well in long sequences with a redetection module 
2 Struck only estimates the location of target and does not handle scale variation
3 Sparse representations are effectivemodels to account for appearance change (e.g., occlusion).
4 Local sparse representations are more effective than the ones with holistic sparse
5 It indicates the alignmentpooling technique adopted by ASLA is more robust to misalignments and background clutters.
6 When an object moves fast, dense sampling based trackers (e.g., Struck, TLD and CXT) perform much better than others
7 On the OCC subset, the Struck, SCM, TLD, LSK and ASLA methods outperform others. The results suggest that structured learning and local sparse representations are effective in dealing with occlusions.
8 On the SV subset,ASLA, SCM and Struck perform best. The results show that
trackers with affine motion models (e.g., ASLA and SCM) often handle scale variation better than others that are designed to account for only translational motion with a few exceptions such as Struck
9 The performance of TLD, CXT, DFT and LOT decreases with the increase of
initialization scale. This indicates these trackers are more sensitive to background clutters. 
10 On the other hand, some trackers perform well or even better when the initial bounding box is enlarged, such as Struck, OAB, SemiT, and BSBT. This indicates that the Haar-like features are somewhat robust to background clutters due to the summation operations when computing features. Overall, Struck is less sensitive to scale variation than other well-performing methods.
11 Some trackers perform better when the scale factor is smaller, such as L1APG, MTT, LOT and CPF