Triplet Network Based on Dynamic Feature Attention for Object Tracking

Fig. 6. Experimental results of different attribute sequences based on OTB100 dataset. (a1)(a2) Background clutters; (b1)(b2) deformation; (c1) (c2) fast motion; (d1) (d2) illumination variation; (e1)(e2) in-plane rotation; (f1)(f2) low resolution; (g1)(g2) motion blur; (h1) (h2) occlusion; (i1) (i2) out-of-plane rotation; (j1)(j2) out-of-view; (k1)(k2) scale variation

Download full size

Fig. 7. Actual tracking results of proposed algorithm and comparison algorithms in video sequences with different attributes. (a) Basketball; (b) Bird1; (c) Board; (d) Ironman; (e) Soccer

Download full size

Table 1. Positive and negative sample allocation of training data in video sequence with DFA-TriNet

View table

Table 1. Positive and negative sample allocation of training data in video sequence with DFA-TriNet

Sample type	Initial frame template image（within 100 frames）	Update template image（within 5 frames）
Positive sample	Same	Same
Negative sample （equally distributed）	Same	Different
	Different	Same
	Different	Different

Table 2. Experimental results for updating parameter based on VOT2018 dataset
View table
Table 2. Experimental results for updating parameter based on VOT2018 dataset
k 1 2 3 4 5 6 7 8 9 10
EAO 0.452 0.466 0.463 0.462 0.469 0.468 0.465 0.464 0.465 0.462

Table 3. Results of ablation experiment based on VOT2018 dataset
View table
Table 3. Results of ablation experiment based on VOT2018 dataset
Index SiamRPN++ SiamRPN+++ BA SiamRPN+++ BA+UB SiamRPN+++ BA+UB+DFA
EAO 0.414 0.448 0.459 0.469
Tracking speed /（frame· $s^{- 1}$ ） 35 44 38 37

Table 4. Experimental results based on VOT2018 dataset

View table

Table 4. Experimental results based on VOT2018 dataset

Index	DaSiamRPN	UPDT	SiamRPN	ATOM	UpdateNet	SiamRPN++	SiamBAN	SiamAttn	Proposed
EAO	0.326	0.378	0.383	0.401	0.403	0.414	0.452	0.470	0.469
Accuracy	0.569	0.536	0.586	0.590	0.583	0.600	0.597	0.630	0.614
Robustness	0.337	0.184	0.276	0.203	0.225	0.234	0.178	0.160	0.122
Tracking speed /（frame· $s^{- 1}$ ）	59	0.4	38	30	55	35	44	33	37

Tools

Get Citation

Copy Citation Text

Zishuo Zhang, Yong Song, Xin Yang, Yufei Zhao, Ya Zhou. Triplet Network Based on Dynamic Feature Attention for Object Tracking[J]. Acta Optica Sinica, 2022, 42(15): 1515001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites