Object Tracking via Dual Linear Structured SVM and Explicit Feature Map

Jifeng Ning, Jimei Yang, Shaojie Jiang, Lei Zhang and Ming-Hsuan Yang

Abstract: Structured support vector machine (SSVM) based methods have demonstrated encouraging performance in recent object tracking benchmarks. However, the complex and expensive optimization limits their deployment in real-world applications. In this paper, we present a simple yet efficient dual linear SSVM (DLSSVM) algorithm to enable fast learning and execution during tracking. By analyzing the dual variables, we propose a primal classifier update formula where the learning step size is computed in closed form. This online learning method significantly improves the robustness of the proposed linear SSVM with lower computational cost. Second, we approximate the intersection kernel for feature representations with an explicit feature map to further improve tracking performance. Finally, we extend the proposed DLSSVM tracker with multi-scale estimation to address the ``drift" problem. Experimental results on large benchmark datasets with 50 and 100 video sequences show that the proposed DLSSVM tracking algorithm achieves state-of-the-art performance.

1. Publication

Figure 1 the proposed DLSSVM tracker.

Jifeng Ning, Jimei Yang, Shaojie Jiang, Lei Zhang and Ming-Hsuan Yang. Object Tracking via Dual Linear Structured SVM and Explicit Feature Map, 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 4266-4274, Las Vegas, USA, 2016. [pdf (605KB)], [matlab sources codes (544kb)]

2. Experimental results

2.1 Analysis of Proposed DLSSVM and Related SSVM Trackers

Table 1. Characteristics of SSVM trackers. NU means no unary representation for the features
SSVM trackers	closed form solution	kernel type	feature type	feature dimensions	high dimension feature	non-linear decesion	Discriminative classifier
SSG	no	linear	image feature	1600	yes	no	explicit
Struck	yes	Gaussian	Haar-like	192	no	yes	implicit
Linear-Struck-NU	yes	linear	image feature	1600	yes	no	explicit
Linear-Struck	yes	linear	image feature	6400	yes	yes	explicit
DLSSVM-NU	yes	linear	image feature	1600	yes	no	explicit
DLSSVM	yes	linear	image feature	6400	yes	yes	explicit

Table 2. Experimental comparisons of the proposed DLSSVM and related trackers with different parameters settings: B50, B100 and B500 mean the budgets of support vectors are 50, 100 and 500 respectively. The entries in red indicate the best results and the ones in blue indicate the second best.
	SSVM trackers	OPE		TRE		SRE		Mean FPS
		precision (20 pixels)	success (AUC)	precision (20 pixels)	success (AUC)	precision (20 pixels)	success (AUC)
	DLSSVM-NU	0.794	0.557	0.810	0.581	0.724	0.508	28.88
	DLSSVM-B50	0.828	0.587	0.846	0.606	0.780	0.543	10.10
	DLSSVM-B100	0.829	0.589	0.856	0.610	0.783	0.545	10.22
	DLSSVM-B500	0.826	0.588	0.852	0.609	0.787	0.548	10.37
	Scale-DLSSVM	0.861	0.608	0.857	0.615	0.811	0.565	5.40
	SSG	0.608	0.443	0.665	0.486	0.584	0.424	46.13
	Struck	0.656	0.474	0.707	0.514	0.634	0.449	0.90
	Linear-Struck-NU	0.703	0.506	0.751	0.540	0.655	0.462	1.46
	Linear-Struck	0.792	0.556	0.824	0.589	0.736	0.515	1.20

2.2 Comparisons with State-of-the-Art Trackers

Figure 2. Average precision plot and success plot for the OPE on the TB50 [1] dataset.

Figure 3. Average precision plot and success plot (bottom row) for the OPE on the TB100 [2] dataset.

2.3 Download the experimental data

DLSSVM and Scale-DLSSVM on OTB-50 and OTB-100 (28MB)

3. Matlab source codes

DLSSVM and Scale-DLSSVM (544kb)

References

[1] Y. Wu, J. Lim, and M.-H. Yang. Online object tracking: A benchmark. In CVPR, 2013.

[2] Y. Wu, J. Lim, and M.-H. Yang. Object tracking benchmark. PAMI, 37(9):1834-1848, 2015.