IEEE Signal Processing Magazine - January 2018 - 85

to their internal intrinsic characteristics
(e.g., living creatures such as cats generally have more deformable appearances
than man-made objects such as vehicles)
OD
or external capturing conditions such as
viewing distances or angles (e.g., deformable objects may appear somewhat rigid
at a distance, and even rigid objects may
(a)
exhibit variations under different viewing
angles). Usually, OD algorithms output
thousands of object proposals or hypotheses as shown in Figure 1(a), which can
SOD
benefit a wide range of computer vision
tasks like weakly supervised learning [3]
and object tracking [4].
SOD [5], [6] is another direction in
(b)
object detection, which aims at mimicking the visual attention mechanism to
highlight objects that draw our attention from each given image [91]. This is
inspired by the human visual attention
COD
system, which can guide humans to pay
special attention to a few informative
Motobike
Car
image regions that are naturally distinct
(the bottom-up saliency) or related to
(c)
certain objects categories determined
by cognitive phenomena like knowl- FIGURE 1. The three research directions in object detection: (a) OD, (b) SOD, and (c) COD.
edge, expectations, reward, and specific tasks (the top-down saliency) [7].
Similar to OD, bottom-up SOD suffers from the challenges that
For addressing the challenging problems in object detection,
come from large appearance variation in unconstrained object
intensive works have been proposed to design better handcrafted
categories, whereas top-down SOD faces the challenge of how
features (such as HOG and SIFT) and put forward sophisticated
to effectively associate the desired visual stimulus (usually at a
object detection frameworks to incorporate the extracted features
semantic level) and the corresponding regions in visual scenes.
with the carefully designed classifiers (e.g., random forest and
Usually, SOD algorithms output a limited number of object
AdaBoost) in the stage of the whole development of object detecregions based on the obtained saliency maps, as shown in
tion. Convolutional neural networks (CNNs) were first applied
Figure 1(b). They can also benefit a wide range of computer
to object detection in 2004 [14] and have been widely used since
vision tasks like image retrieval [8] and object segmentation [9].
2013 [15]. The work on region-based CNNs (RCNNs) in [10] led
The third direction of object detection is COD [10], [11]. Difto significant breakthroughs in 2014. It made the earliest efforts
ferent from OD, COD aims at detecting multiple predefined
to describe an object detection system by using multilayer conobject categories from each given image. It needs to not only
volutional networks to extract highly discriminative yet invariant
identify the image regions that may contain the objects of interest
feature representation.
but also recognize the specific object category of each detected
This work has achieved a significant improvement of more
image region. Compared with SOD, COD has a totally differthan 50% mean average precision (mAP) when compared to the
ent motivation, i.e., it moves toward solving a pure computational
best methods at that time, which were based on the handcraftproblem without any understanding of the function in the human
ed image features on the commonly used PASCAL detection
visual system, e.g., visual attention. Usually, COD is converted
benchmark [16]. Since then, several advanced deep-learningto a multiclass classification problem, where discriminative clasbased techniques [17]-[20] have been proposed for high-quality
sification functions are trained to separate the extracted image
object detection, which covers all three related areas of OD,
regions in the corresponding feature domain. The main challenge
SOD, and COD. To this end, this article presents a comprehenin COD is how to deal with the intraclass appearance variation
sive survey of recent state-of-the-art approaches.
and interclass appearance similarity. As shown in Figure 1(c),
This article mainly has four motivations:
COD approaches usually output multiple image regions assigned
1) Object detection, including OD, SOD, and COD, is a funwith the identified object category. COD can be applied to
damental yet challenging problem of computer vision. The
computer vision tasks like scene parsing [12] and human action
existing survey papers focus on each individual topic only,
recognition [13].
without discussing the close relationship.
IEEE SIGNAL PROCESSING MAGAZINE

|

January 2018

|

85
Table of Contents for the Digital Edition of IEEE Signal Processing Magazine - January 2018
