![](https://csdnimg.cn/release/download_crawler_static/15117468/bg4.jpg)
372 X. Lu et al.
displacement between samples for precise localization. Two-stage deep trackers
benefit little from the recent advance of ROI pooling [29], which cannot highlight
the difference between highly spatially correlated samples.
One-Stage Tracking. The one-stage tracking framework takes the whole search
area as input and directly outputs a response map through a learned regres-
sor, which learns a mapping between input features and soft labels generated
by a Gaussian function. One representative category of one-stage trackers are
based on discriminative correlation filters [30], which regress all the circularly
shifted versions of input image into soft labels. By computing the correlation
as an element-wise product in the Fourier domain, DCFs trackers achieve the
fastest speed thus far. Numerous extensions include KCF [31], LCT [32,33],
MCF [34], MCPF [35] and BACF [14]. With the use of deep features, DCFs
trackers, such as DeepSRDCF [1], HDT [2], HCFT [3], C-COT [4]andECO[5],
have shown superior performance on benchmark datasets. In [3], Ma et al. pro-
pose to learn multiple DCFs over different convolutional layers and empirically
fuse output correlation maps to locate target objects. A similar idea is exploited
in [4] to combine multiple response maps. In [5], Danelljan et al. reduce feature
channels to accelerate learning correlation filters. Despite the top performance,
DCFs trackers independently extract deep features to learn and update corre-
lation filters. In the deep learning era, DCFs trackers can hardly benefit from
end-to-end training. The other representative category of one-stage trackers are
based on convolutional regression networks. The recent FCNT [6], STCT [7], and
CREST [8] trackers belong to this category. The FCNT makes the first effort
to learn regression networks over two CNN layers. The output response maps
from different layers are switched according to their confidence to locate tar-
get objects. Ensemble learning is exploited in the STCT to select CNN feature
channels. CREST [8] learns a base network as well as a residual network on a
single convolutional layer. The output maps of the base and residual networks
are fused to infer target positions. We note that current deep regression trackers
do not perform as well as DCFs trackers. We identify the main bottleneck as
the data imbalance issue in regression learning. By balancing the importance
of training data, the performance of one-stage deep regression trackers can be
significantly improved over state-of-the-art DCFs trackers.
Data Imbalance. The data imbalance issue has been extensively studied in the
learning community [11,36,37]. Helpful solutions involve data re-sampling [38–
40], and cost-sensitive loss [16,41–43]. For visual tracking, Li et al. [44]usea
temporal sampling scheme to balance positive and negative samples to facilitate
CNN training. Bertinetto et al. [45] balance the loss of positive and negative
examples in the score map for pre-training the Siamese fully convolution network.
The MDNet [26] tracker shows that it is crucial to mine the hard negative samples
during training classification networks. The recent work [16] on dense object
detection proposes focal loss to decrease the loss from imbalance samples. Despite
the importance, current deep regression trackers [6–
8] pay little attention to
data imbalance. In this work, we propose to utilize shrinkage loss to penalize
easy samples which have little contribution to learning regression networks. The