Frequency Perception Network for Camouflaged Object

Detection
Runmin Cong Mengyao Sun Sanyi Zhang∗
School of Control Science and Institute of Information Science, State Key Laboratory of Information
Engineering, Key Laboratory of Beijing Key Laboratory of Advanced Security (SKLOIS),
Machine Intelligence and System Information Science and Network Institute of Information Engineering,
Control, Ministry of Education, Technology, Chinese Academy of Sciences
Shandong University Beijing Jiaotong University Beijing, China
Jinan, Shandong, China Beijing, China zhangsanyi@[Link]
rmcong@[Link] sunmengyao@[Link]
arXiv:2308.08924v2 [[Link]] 7 Dec 2024

Xiaofei Zhou Wei Zhang Yao Zhao

School of Automation, School of Control Science and Institute of Information Science,
Hangzhou Dianzi University Engineering, Key Laboratory of Beijing Key Laboratory of Advanced
Hangzhou, Zhejiang, China Machine Intelligence and System Information Science and Network
zxforchid@[Link] Control, Ministry of Education, Technology,
Shandong University Beijing Jiaotong University
Jinan, Shandong, China Beijing, China
davidzhang@[Link] yzhao@[Link]

ABSTRACT CCS CONCEPTS

Camouflaged object detection (COD) aims to accurately detect ob- • Computing methodologies → Scene understanding.
jects hidden in the surrounding environment. However, the exist-
ing COD methods mainly locate camouflaged objects in the RGB KEYWORDS
domain, their performance has not been fully exploited in many Camouflaged object detection, Frequency perception, Coarse posi-
challenging scenarios. Considering that the features of the cam- tioning stage, Fine localization stage.
ouflaged object and the background are more discriminative in
the frequency domain, we propose a novel learnable and separa- ACM Reference Format:
ble frequency perception mechanism driven by the semantic hi- Runmin Cong, Mengyao Sun, Sanyi Zhang, Xiaofei Zhou, Wei Zhang,
and Yao Zhao. 2023. Frequency Perception Network for Camouflaged Object
erarchy in the frequency domain. Our entire network adopts a
Detection. In Proceedings of the 31st ACM International Conference on Multi-
two-stage model, including a frequency-guided coarse localiza- media (MM ’23), October 29-November 3, 2023, Ottawa, ON, Canada. ACM,
tion stage and a detail-preserving fine localization stage. With New York, NY, USA, 11 pages. [Link]
the multi-level features extracted by the backbone, we design a
flexible frequency perception module based on octave convolu-
tion for coarse positioning. Then, we design the correction fusion
1 INTRODUCTION
module to step-by-step integrate the high-level features through In nature, animals use camouflage to blend in with their surround-
the prior-guided correction and cross-layer feature channel asso- ings to avoid detection by predators. The camouflaged object detec-
ciation, and finally combine them with the shallow features to tion (COD) task aims to allow computers to automatically recognize
achieve the detailed correction of the camouflaged objects. Com- these camouflaged objects that blend in with the background, which
pared with the currently existing models, our proposed method can be used in numerous downstream applications, including med-
achieves competitive performance in three popular benchmark ical segmentation [11, 14, 21, 28], unconstrained face recognition
datasets both qualitatively and quantitatively. The code will be [3], and recreational art [15, 23]. However, the COD task is very
released at [Link] challenging due to the low contrast properties between the cam-
ouflaged object and the background. Furthermore, camouflaged
∗ Corresponding author objects may have multiple appearances, including shapes, sizes,
and textures, which further increases the difficulty of detection.
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed At the beginning of the research, the COD task was also re-
for profit or commercial advantage and that copies bear this notice and the full citation garded as a low-contrast special case of the salient object detection
on the first page. Copyrights for components of this work owned by others than the (SOD) task, but simple SOD model [5–10, 12, 13, 30, 62] retrain-
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specific permission ing cannot obtain satisfactory COD results, and usually requires
and/or a fee. Request permissions from permissions@[Link]. some special positioning design to find camouflaged objects. Re-
MM ’23, October 29-November 3, 2023, Ottawa, ON, Canada cently, with the development of deep learning [22, 27, 46, 61, 64],
© 2023 Copyright held by the owner/author(s). Publication rights licensed to ACM.
ACM ISBN 979-8-4007-0108-5/23/10. . . $15.00 many customized networks for COD tasks have gradually emerged
[Link] [2, 20, 41]. However, current solutions still struggle in challenging
MM ’23, October 29-November 3, 2023, Ottawa, ON, Canada Runmin Cong et al.

and cross-layer feature channel association. Finally, the shallow

high-resolution features are further introduced to refine and modify
the boundaries of camouflaged objects and generate the final COD
result.
The main contributions are summarized as follows:

• We propose a novel two-stage framework to deeply exploit

the advantages of RGB and frequency domains for camou-
flaged object detection in an end-to-end manner. The pro-
posed network achieves competitive performance on three
popular benchmark datasets (i.e., COD10K, CHAMELEON,
and CAMO).

• A novel fully frequency-perception module is designed to

enhance the ability to distinguish camouflaged objects from
Figure 1: Three challenging camouflaged object detection
backgrounds by automatically learning high-frequency and
(COD) scenarios from top to down are with indefinable bound-
low-frequency features, thereby achieving coarse localiza-
aries, multiple objects, and occluded objects, respectively.
tion of camouflaged objects.
The images from left to right are (a) Input image, (b) GT, (c)
Ours, (d) SINet-V2 [16], (e) LSR [42]. • We design a progressive refinement mechanism to obtain the
final refined camouflaged object detection results through
prior-guided correction, cross-layer feature channel associa-
situations, such as multiple camouflaged objects, uncertain or fuzzy tion, and shallow high-resolution boundary refinement.
object boundaries, and occlusion, as shown in Figure 1. In general,
these methods mainly design modules in the RGB color domain to
detect camouflaged objects, and complete the initial positioning of 2 RELATED WORK
camouflaged objects by looking for areas with inconsistent infor- The COD task aims to localize objects that have a similar appear-
mation such as textures (called breakthrough points). However, the ance to the background, which makes it extremely challenging.
concealment and confusion of the camouflaged objects itself make Early methods employed hand-crafted low-level features to achieve
this process very difficult. In the image frequency domain analysis, this goal, such as color [29], expectation-maximization statistics
the high-frequency and low-frequency component information in [39], convex intensity [52], optical flow [26], and texture [1, 31].
the frequency domain describes the details and contour character- However, due to the imperceptible differences between objects and
istics of the image in a more targeted manner, which can be used to backgrounds in complex environments, and the limited expressive
improve the accuracy of the initial positioning. Inspired by this, we power of hand-crafted features, they do not perform satisfactorily.
propose a Frequency Perception Network (FPNet) that employs a Recently CNN-based methods [20, 33, 45] have achieved signifi-
two-stage strategy of search and recognition to detect camouflaged cant success in the COD task. In general, CNN-based methods often
objects, taking full advantage of RGB and frequency cues. employ one or more of the following strategies, such as two-stage
On the one hand, the main purpose of the frequency-guided strategy [2, 20], multi-task learning strategy [42], and incorporat-
coarse positioning stage is to use the frequency domain features to ing other guiding cues such as frequency [24]. For instance, Fan et
find the breakthrough points of the camouflaged object position. al. [20] proposed a two-stage process named SINet, which repre-
We first adopt the transformer backbone to extract multi-level fea- sents the new state-of-the-art in existing COD datasets and created
tures of the input RGB image. Subsequently, in order to realize the the largest COD10K dataset with 10K images. Mei et al. [45] imi-
extraction of frequency domain features, we introduce a frequency- tated the predator-prey process in nature and developed a two-stage
perception module to decompose color features into high-frequency bionic framework called PFNet.
and low-frequency components. Among them, the high-frequency In terms of frequency domain studies, Gueguen et al. [24] di-
features describe texture features or rapidly changing parts, while rectly used the Discrete Cosine Transform (DCT) coefficients of
the low-frequency features can outline the overall contour of the the image as input to CNN for subsequent visual tasks. Ehrlich et
image. Considering that both texture and contour are important for al. [17] presented a general conversion algorithm for transforming
camouflaged object localization, we fuse them as a complete repre- spatial domain networks to the frequency domain. Interestingly,
sentation of frequency domain information. In addition, a neighbor both of these works delve deep into the frequency domain trans-
interaction mechanism is also employed to combine different levels formation of the image JPEG compression process. Subsequently,
of frequency-aware features, thereby achieving coarse detection Zhong et al. [65] modeled the interaction between the frequency
and localization of camouflaged objects. On the other hand, the domain and the RGB domain, introducing the frequency domain as
detail-preserving fine localization stage focuses on progressively an additional cue to better detect camouflaged objects from back-
prior-guided correction and fusion across layers, thereby gener- grounds. Unlike these methods, on the one hand, we use octave
ating the final finely camouflaged object masks. Specifically, we convolution to realize the online learning of frequency domain
design the correction fusion module to achieve the cross-layer high- features, instead of offline extraction methods (e.g., DCT); on the
level feature interaction by integrating the prior-guided correction other hand, frequency domain features are mainly used for coarse
Frequency Perception Network for Camouflaged Object Detection MM ’23, October 29-November 3, 2023, Ottawa, ON, Canada

Figure 2: The overview of our proposed two-stage network FPNet. The input image is first extracted with multi-level features
by a PVT encoder. In the frequency-guided coarse localization stage, we use FPM for frequency-domain feature extraction and
generate the coarse COD map 𝑆 1 . Then, in the detail-preserving fine localization stage, the CFM is used to achieve progressively
prior-guided correction and fusion across high-level layers. Finally, the first-level high-resolution features are further introduced
to refine the boundaries of camouflaged objects and generate the final result 𝑆𝑜𝑢𝑡𝑝𝑢𝑡 .

positioning in the first stage, that is, by making full use of high- end, in this paper, we propose a Frequency Perception Network
frequency and low-frequency information to find the breakthrough (FPNet) for camouflaged object detection, as shown in Figure 2,
point of camouflaged object positioning in the frequency domain. including a feature extraction backbone, a frequency-guided coarse
In addition, some methods [37, 47, 51] also try to combine edge localization stage, and a detail-preserving fine localization stage.
detection to extract more precise edges, thereby improving the Given an input image 𝐼 ∈ R𝐻 ×𝑊 ×3 , for the feature extraction
accuracy of COD. It is worth mentioning that in order to exploit the backbone, we adopt the Pyramid Vision Transformer (PVT) [53]
power of the Transformer model in the COD task, many Transformer- as the encoder to generate features of different levels, denoted as
based methods have emerged. For example, Yang et al. [58] proposed 𝑋𝑖 (𝑖 = {1, 2, 3, 4}). Each feature map serves a different purpose.
to incorporate Bayesian learning into Transformer-based reasoning The first-level feature map 𝑋 1 includes rich detailed information
to achieve the COD task. The T2Net proposed by Mao et al. [43] in about the camouflaged object, whereas the deeper-level features
2021 used a Swin-Transformer as the backbone network, surpassing (𝑋 2 , 𝑋 3 , 𝑋 4 ) contain higher-level semantic information. With the
all CNN-based approaches at that time. pyramid backbone features, in the frequency-guided coarse local-
ization stage, we first use a frequency-perception module (FPM)
for frequency-domain feature extraction on high-level features and
3 OUR APPROACH then adopt the neighborhood connection decoder for feature fu-
3.1 Overview sion decoding to obtain the coarse COD map 𝑆 1 . Whereafter, in the
Our goal is to exploit and fuse the inherent advantages of the RGB detail-preserving fine localization stage, with the guidance of coarse
and frequency domains to enhance the discrimination ability to COD map 𝑆 1 , the high-level features are embedded into the correc-
discover camouflaged objects in the complex background. To that tion fusion module (CFM) to progressively achieve prior-guided
MM ’23, October 29-November 3, 2023, Ottawa, ON, Canada Runmin Cong et al.

Figure 3: Illustration of frequency-perception module (FPM).

Two branches are for high-frequency and low-frequency in-
formation learning, respectively. Figure 4: The schematic illustration of the correction fusion
module (CFM). CFM contains two parts, i.e., prior-guided
correction and channel-wise correlation modeling.
correction and fusion across layers. Finally, a receptive field block
(RFB) with spatial attention mechanism (SAM) is used for low-level
high-resolution feature optimization and combined with the CFM domain information:
module output to obtain the final COD result 𝑆𝑜𝑢𝑡𝑝𝑢𝑡 . 𝑓𝑖 = Resize(𝑌𝑖𝐻 ) ⊕ Resize(𝑌𝑖𝐿 ), (3)
where Resize means to adjust features to a fixed dimension, and ⊕
3.2 Frequency-guided Coarse Positioning is the element-wise addition.
Inspired by predator hunting systems, frequency information is Then, the Neighbor Connection Decoder (NCD) [16], as shown
more advantageous than RGB appearance features for a specific in the top region (the part above the three FPMs) of Figure 2, is
predator in the wild environment. This point of view has also been adopted to gradually integrate the frequency-domain features of
verified in [65], and then a frequency domain method for cam- the top-three layers, fully utilizing the cross-layer semantic context
ouflaged object detection is proposed. Specifically, this work [65] relationship through the neighbor layer connection, which can be
used offline discrete cosine transform to convert the RGB domain represented as:
information of an image to the frequency domain, but the offline
frequency extraction method limits its flexibility. As described in ′
 𝑓4 = ℊ ↑ (𝑓4 ),


𝑓3′ = 𝑓3 ⊗ ℊ ↑ (𝑓4 )

[4], octave convolution can learn to divide an image into low and (4)
high frequency components in the frequency domain. The low-  𝑓 ′ = cat(𝑓2 ⊗ ℊ ↑ (𝑓 ′ ), cat(𝑓 ′, 𝑓 ′ )),

 2 3 3 4
frequency features correspond to pixel points with gentle intensity where ⊗ is element-wise multiplication, ℊ ↑ (𝑥) denotes an up-
transformations, such as large color blocks, that often represent sampling along with a 3×3 convolution, cat() denotes concatenation
the main part of the object. The high-frequency components, on along with a 3 × 3 convolution, and 𝑓2′ is the output of NCD. After
the other hand, refer to pixels with intense brightness changes, this stage, we use a simple convolution to obtain a coarse mask 𝑆 1
such as the edges of objects in the image. Inspired by this, we that reveals the initial location of the camouflaged object.
propose a frequency-perception module to automatically separate
features into high-frequency and low-frequency parts, and then 3.3 Detail-preserving Fine Localization
form a frequency-domain feature representation of camouflaged
In the previous section, we introduced how to use frequency-domain
objects, the detailed process is shown in Figure 3.
features to achieve coarse localization of camouflaged objects. But
Specifically, we employ octave convolution [4] to automatically
the first stage is more like a process of finding and locating break-
perceive high-frequency and low-frequency information in an end-
through points, the integrity and accuracy of results are still not
to-end manner, enabling online learning of camouflaged object
enough. To this end, we propose a detail-preserving fine localization
detection. The octave convolution can effectively avoid blockiness
mechanism, which not only achieves a progressive fusion of high-
caused by the DCT and utilize the advantage of the computational
level features through prior correction and channel association but
speed of GPUs. In addition, it can be easily plugged into arbitrary
also considers high-resolution features to refine the boundaries of
networks. The detailed process of output of the octave convolution
camouflaged objects, as shown in Figure 2.
𝑌𝑖 = {𝑌𝑖𝐻 , 𝑌𝑖𝐿 } could be described in the following:
To achieve the above goals, we first design a correction fusion
𝑌𝑖𝐻 = 𝐹 (𝑋𝑖𝐻 ;𝑊 𝐻 →𝐻 ) + Upsample(𝐹 (𝑋𝑖𝐿 ;𝑊 𝐿→𝐻 ), 2), (1) module (CFM), which effectively fuses adjacent layer features and
a coarse camouflaged mask to produce fine output. The module
includes three inputs: the current and previous layer features 𝑋𝑖
𝑌𝑖𝐿 = 𝐹 (𝑋𝑖𝐿 ;𝑊 𝐿→𝐿 ) + 𝐹 (pool(𝑋𝑖𝐻 , 2);𝑊 𝐻 →𝐿 ), (2)
and 𝑋𝑖+1 , and the coarse mask 𝑆𝑔 = {𝑆 1, 𝑆 2 }. In addition, we first re-
where 𝐹 (𝑋 ;𝑊 ) denotes a convolution with the learnable parame- duce the number of input feature channels to 64, denoted as 𝐹𝑖 and
ters of 𝑊 , pool(𝑋, 𝑘) is an average pooling operation with kernel 𝐹𝑖+1 , which helps to improve computational efficiency while still
size of 𝑘 × 𝑘, and Upsample(𝑋, 𝑠) is an up-sampling operation by a retaining relevant information for detection. As shown in Figure
factor of 𝑠 via nearest interpolation. 4, our CFM consists of two parts. In order to make full use of the
Considering that both high-frequency texture attribute and low- existing prior guidance map 𝑆𝑔 , we purify the features of the previ-
frequency contour attribute are important for camouflaged object ous layer and select the features most related to the camouflaged
localization, we fuse them as a complete representation of frequency features to participate in the subsequent cross-layer interaction.
Frequency Perception Network for Camouflaged Object Detection MM ’23, October 29-November 3, 2023, Ottawa, ON, Canada

Mathematically, the feature map 𝐹𝑖+1 is first multiplied with the 4 EXPERIMENT
′
coarse mask 𝑆𝑔 to obtain the output features 𝑓𝑖+1 : 4.1 Experimental Settings
′
𝑓𝑖+1 = Upsample(𝐹𝑖+1 ⊙ 𝑆𝑔 ), (5) Datasets. We conduct experiments and evaluate our proposed
method on three popular benchmark datasets, i.e., CHAMELEON
where ⊙ denotes element-wise multiplication, and Upsample is the [49], CAMO [33], and COD10K [20]. CHAMELEON [49] dataset has
upsampling operation. This prior-guided correction is particularly 76 images. CAMO [33] contains 1,250 camouflaged images covering
beneficial in scenarios where the object is difficult to discern from different categories, which are divided into 1,000 training images
its surroundings. and 250 testing images, respectively. As the largest benchmark
It is well known that high-level features possess very rich channel- dataset currently, COD10K [20] includes 5,066 images in total, 3,040
aware cues. In order to achieve more sufficient cross-layer feature images are chosen for training and 2,026 images are used for testing.
interaction and effectively transfer the high-level information of There are five concealed super-classes (i.e., terrestrial, atmobios,
the previous layer to the current layer, we design the channel-level aquatic, amphibian, other) and 69 sub-classes. And the pixel-level
association modeling. We perform channel attention by taking the ground-truth annotations of each image in these three datasets
′
inner product between each pixel point on 𝐹𝑖 and 𝑓𝑖+1 , which calcu- are provided. Besides, for a fair comparison, we follow the same
lates the similarity between different feature maps in the channel training strategy of previous works [65], our training set includes
dimension of the same pixel. To further reduce computational com- 3,040 images from COD10K datasets and 1,000 images from the
plexity, we also employ a 3 × 3 convolution that creates a bottleneck CAMO dataset.
structure, thereby compressing the number of output channels. This Evaluation Metrics. We use four widely used and standard metrics
process can be described as: to evaluate the proposed method, i.e., structure-measure (𝑆𝛼 ) [18],
′ 𝑇
mean E-measure (𝐸𝜙 ) [19], weighted F-measure (𝐹 𝛽𝜔 ) [44], and
𝐴 = conv(𝐹𝑖 ⊗ (𝑓𝑖+1 ) ), (6)
mean absolute error (𝑀𝐴𝐸) [34, 35, 63]. Overall, a better COD
where ⊗ is the matrix multiplication. Then, we learn two weight method has larger 𝑆𝛼 , 𝐸𝜙 , and 𝐹 𝛽𝜔 scores, but a smaller 𝑀𝐴𝐸 score.
maps, 𝛼 and 𝛽, by using two 3 × 3 convolution operations on the Implementation Details. In this paper, we propose a frequency-
features 𝐴. They are further used in the correction of the features of perception network (FPNet) to address the challenge of camouflaged
the current layer 𝐹𝑖 in a modulation manner. In this way, the final object detection by incorporating both RGB and frequency domains.
cross-level fusion features can be generated through the residual Specifically, a frequency-perception module is proposed to automat-
processing: ically separate frequency information leading the model to a good
′
𝑓𝑖𝑜𝑢𝑡 = 𝑓𝑖+1 + conv(𝐹𝑖 ) ∗ 𝛼 + 𝛽. (7) coarse mask at the first stage. Then, a detail-preserving fine localiza-
tion module equipped with a correction fusion module is explored to
In addition to the above-mentioned prior correction and channel- refine the coarse prediction map. Comprehensive comparisons and
wise association modeling on the high-level features, we also make ablation studies on three benchmark COD datasets have validated
full use of the high-resolution information of the first layer to sup- the effectiveness of the proposed FPNet. The proposed method is
plement the detailed information. Specifically, we use the receptive implemented with PyTorch and leverages Pyramid Vision Trans-
field block (RFB) module [38] and the spatial attention module [56] former [53] pre-trained on ImageNet [32] as our backbone network.
on the first-layer features (𝑋 1 ) to enlarge the receptive field and We also implement our network by using the MindSpore Lite tool1 .
highlight the important spatial information of the features, and To update the network parameters, we use the Adam optimizer,
then fuse with the output of the CFM module (𝑓2𝑜𝑢𝑡 ) to generate which is widely used in transformer-based networks [40, 53, 54].
the final prediction map: The initial learning rate is set to 1e-4 and weight decay is adjusted
to 1e-4. Furthermore, we resize the input images to 512 × 512, the
𝑆𝑜𝑢𝑡𝑝𝑢𝑡 = 𝐵𝑐𝑜𝑛𝑣 (𝐵𝑐𝑜𝑛𝑣 (𝑆𝐴𝑀 (𝑅𝐹 𝐵(𝑋 1 )) ⊕ 𝑓2𝑜𝑢𝑡 )), (8)
model is trained with a mini-batch size of 4 for 100 epochs on an
where 𝑅𝐹 𝐵 and 𝑆𝐴𝑀 are the receptive field block and the spatial at- NVIDIA 2080Ti GPU. We augment the training data by applying
tention module, respectively. 𝐵𝑐𝑜𝑛𝑣 represents the 3×3 convolution techniques such as random flipping, random cropping, and so on.
layer along with the batch normalization and ReLU.
4.2 Comparison with State-of-the-art Methods
3.4 Loss Function We conduct a comparison of our proposed method with 12 state-
Following [55], We compute the weighted binary cross-entropy of-the-art mthods, including FPN [36], MaskRCNN [25], CPD [57],
𝜔 ) and IoU loss (L 𝜔 ) on three COD maps (i.e., 𝑆 , 𝑆 ,
loss (L𝐵𝐶𝐸 SINet [20], LSR [42], PraNet [21], C2FNet [50], UGTR [58], PFNet
𝐼𝑜𝑈 1 2
and 𝑆𝑜𝑢𝑡𝑝𝑢𝑡 ) to form our final loss function: [45], ZoomNet [48], SINet-V2 [16], and FreNet [65]. The visualiza-
tion comparisons and quantitative results are shown in Figure 5, and
L𝑡𝑜𝑡𝑎𝑙 = L1 + L2 + L𝑜𝑢𝑡𝑝𝑢𝑡 , (9) Table 1 summarizes the quantitative results of the COD methods
𝜔 𝜔 , ∗ = {1, 2, 𝑜𝑢𝑡𝑝𝑢𝑡 }, L denotes the
on three benchmark datasets.
where L∗ = L𝐵𝐶𝐸 + L𝐼𝑜𝑈 1 Quantitative Evaluation. Table 1 presents a detailed comparison
loss between the coarse prediction map 𝑆 1 and ground truth, L2 of evaluation metrics, we can observe that our proposed model
denotes the loss about the prediction map 𝑆 2 after the first CFM, (FPNet) outperforms all SOTA models on all datasets. For example,
and L𝑜𝑢𝑡𝑝𝑢𝑡 denotes the loss between the final prediction map
𝑆𝑜𝑢𝑡𝑝𝑢𝑡 and ground truth. 1 [Link]
MM ’23, October 29-November 3, 2023, Ottawa, ON, Canada Runmin Cong et al.

Table 1: Comparisons of state-of-the-art methods on COD datasets. The top three results are highlighted in red, green, and blue,
respectively.

COD10K-Test (2026 images) CAMO-Test (250 images) CHAMELEON (76 images)

Methods Year
𝑆𝛼 ↑ 𝐸𝑚𝑒𝑎𝑛 ↑ 𝐹 𝛽𝜔 ↑ 𝑀 ↓ 𝑆𝛼 ↑ 𝐸𝑚𝑒𝑎𝑛 ↑ 𝐹 𝛽𝜔 ↑ 𝑀 ↓ 𝑆𝛼 ↑ 𝐸𝑚𝑒𝑎𝑛 ↑ 𝐹 𝛽𝜔 ↑ 𝑀 ↓
FPN [36] 2017-CVPR 0.697 0.691 0.411 0.075 0.684 0.677 0.483 0.131 0.794 0.783 0.590 0.075
MaskRCNN [25] 2017-ICCV 0.613 0.748 0.402 0.080 0.574 0.715 0.430 0.151 0.643 0.778 0.518 0.099
CPD [57] 2019-CVPR 0.747 0.770 0.508 0.059 0.726 0.729 0.550 0.115 0.853 0.866 0.706 0.052
SINet [20] 2020-CVPR 0.771 0.806 0.551 0.051 0.751 0.771 0.606 0.100 0.869 0.891 0.740 0.044
PraNet [21] 2020-MICCAI 0.789 0.857 0.608 0.047 0.774 0.828 0.680 0.094 0.871 0.924 0.758 0.037
PFNet [45] 2021-CVPR 0.798 0.874 0.646 0.040 0.773 0.829 0.703 0.086 0.878 0.921 0.796 0.034
C2FNet [50] 2021-IJCAI 0.809 0.884 0.662 0.038 0.787 0.840 0.716 0.085 0.892 0.946 0.819 0.030
UGTR [58] 2021-ICCV 0.818 0.850 0.667 0.035 0.785 0.859 0.686 0.086 0.888 0.918 0.796 0.031
LSR [42] 2021-CVPR 0.767 0.861 0.611 0.045 0.712 0.791 0.583 0.104 0.846 0.913 0.767 0.046
SINet-V2 [16] 2022-TPAMI 0.815 0.886 0.664 0.036 0.809 0.864 0.729 0.073 0.888 0.940 0.797 0.029
FreNet [65] 2022-CVPR 0.833 0.907 0.711 0.033 0.828 0.884 0.747 0.069 0.894 0.950 0.819 0.030
ZoomNet [48] 2022-CVPR 0.838 0.911 0.729 0.029 0.820 0.892 0.752 0.066 0.902 0.958 0.845 0.023
FPNet (ours) 2023 0.850 0.913 0.748 0.029 0.852 0.905 0.806 0.056 0.914 0.961 0.856 0.022

our FPNet achieves obvious performance gains over other state-

of-the-art ones on the CAMO-Test dataset. According to Table 1,
our proposed FPNet achieves the best weighted F-measure (𝐹 𝛽𝜔 )
score of 0.806 on the CAMO-Test dataset, and the MAE score out-
performs the second-best method ZoomNet [48] by 15.2%. More-
over, the proposed FPNet outperforms ZoomNet [48] by an obvious
margin in terms of the 𝐹 𝛽𝜔 on all three datasets. For example, com-
pared with the second-best method, the percentage gain of the
𝐹 𝛽𝜔 reach 2.6%, 7.2%, and 1.3% on the COD10K-Test, CAMO-Test
and CHAMELEON datasets, respectively. While we observe the
frequency-guided method FreNet [65], it can achieve better perfor-
mance than most state-of-the-art methods. However, our proposed
FPNet outperforms FreNet comprehensively in terms of all eval-
uation metrics, indicating that the proposed learnable frequency- Figure 5: Qualitative results of our proposed FPNet model
guided solution is superior in discerning discriminative cues of and some state-of-the-art COD methods. The images from
camouflaged objects. left to right are (a) Input image, (b) GT, (c) Ours, (d) FreNet
Qualitative Evaluation. As shown in Figure 5, whether the cam- [65], (e) SINet-V2 [16], (f) PFNet [45], (g) LSR [42], and (h)
ouflaged object in the image is terrestrial or aquatic, or a camou- PraNet [21].
flaged human, the proposed FPNet method is capable of accurately
predicting the region of the camouflaged object. When the camou-
flaged object is extremely similar to the background, as illustrated displayed in Figure 1), such as indefinable boundaries, multiple ob-
in the first row in Figure 5, other SOTA methods fail to accurately jects and occluded objects. The impressive prediction results further
distinguish the camouflaged object, especially on the edge regions. highlight the usefulness of the frequency-perception mechanism
By contrast, the proposed FPNet, benefiting from a frequency-aware which connects RGB-aware and frequency-aware clues together
learning mechanism, can clearly predict the mask of objects with to arrive at a unified solution that can adaptly address challenging
clear and sharp boundaries. When tackling complex background scenarios.
interference, including the salient but non-camouflaged objects
(see the third row of Figure 5), our proposed FPNet is capable of 4.3 Ablation Studies
effectively separating the camouflaged object from the background,
4.3.1 The effectiveness of each module of FPNet. To verify
with a more complete and clear structure description ability. For the
the effectiveness of the proposed network, we separate FPNet into
approximate appearance of similar objects, as shown in the fourth
a series of ablation parts, i.e., frequency perception module, high-
row of Figure 5, the camouflaged human face is hard to distinguish
resolution preserving, and correction fusion module, where ‘base-
from other pineapples. Most methods fail to recognize it, but our
line’ is the PVT backbone for camouflaged object detection. The
proposed FPNet can discern it clearly. Additionally, our proposed
comparison results are shown in Table 2.
FPNet is also effective in detecting some challenging situations (as
Effectiveness of Frequency Perception Module. The proposed
frequency perception module incorporates Octave convolution [4]
Frequency Perception Network for Camouflaged Object Detection MM ’23, October 29-November 3, 2023, Ottawa, ON, Canada

Table 2: Quantitative results of ablation studies on the

COD10k-Test dataset. First-stage and Second-stage mean the
Frequency-guided Coarse Positioning and Detail-preserving
Fine Localization respectively. HRP denotes High-res Pre-
serving.
Figure 6: Comparison results of different feature sets output
1st-stage 2nd-stage COD10k-Test (2026 images) from frequency perception module. (a) Input image. (b) GT.
Baseline
FPM HRP CFM 𝑆𝛼 ↑ 𝐸𝑚𝑒𝑎𝑛 ↑ 𝐹 𝛽𝜔 ↑ 𝑀 ↓ (c) Prediction result with our frequency fusing mechanism.
✓ 0.835 0.899 0.701 0.032 (d) Prediction with only high-frequency features. (e) Result
✓ ✓ 0.844 0.908 0.728 0.031
✓ ✓ ✓ 0.849 0.911 0.739 0.030 with only low-frequency features.
✓ ✓ ✓ ✓ 0.850 0.913 0.748 0.029

Effectiveness of Correction Fusion Module. Though the high-

resolution preserving mechanism for detail refining achieves good
performance, the coarse camouflaged mask from the frequency-
that minimizes spatial redundancy and automatically learns high- guided stage is still not effectively exploited enough. Thus, we
frequency and low-frequency features. As shown in Table 2, if propose a correction fusion module to further improve the quality
we add the frequency perception module (i.e., baseline+FPM), all of the camouflaged mask by completely mining the ability of the
metrics can obtain performance gains compared with the PVT- coarse map and the neighbor features. Specifically, we implement
alone without the octave convolution. The good performance lies the CFM on the detail-preserving fine localization stage, the re-
that FPM learns rich frequency-aware information especially the sults are shown in the last row of Table 2. While we update the
high-frequency clues that are useful for camouflaged object coarse detail-preserving with the CFM mechanism, all metric scores can
positioning. The other advantage of the FPM lies in that it is auto- be further improved especially in terms of the 𝑆𝛼 and 𝐹 𝛽𝜔 scores.
matically online learning without any other extra offline operations.
The good target detection ability indicates that CFM plays an essen-
Thus, the flexibility and the high performance of the FPM make it
tial role in improving the detection performance of camouflaged
suitable for accurately detecting camouflaged objects in real-world
objects. The main reason lies that CFM takes the prior coarse cam-
scenes. And FPM can also be easily integrated into other frame-
ouflaged mask and the neighbor layer interaction into account.
works to assist in distinguishing the obscure boundaries of objects
First, the prior coarse prediction mask can provide us with an ac-
that are similar to the background.
curate region of the highlighted camouflaged objects which can
Effectiveness of High-resolution Preserving Module. Although
extract object-centric related features well. Then, the channel-wise
The first frequency-guided coarse positioning stage (i.e., PVT+FPM)
correlation correlates and combines neighbor layers to enhance the
has achieved good target prediction maps, the object boundaries
object representation which is more distinguishable for perceiving
are still unsatisfactory. Thus, we adopt the high-resolution preserv-
the camouflaged objects. Since CFM learns the channel correlation
ing mechanism for further detail refining. As shown in Table 2,
between adjacent features to obtain learnable weight maps and
we conduct the detail-preserving fine localization stage without
adjust original features, the dynamic mechanism achieves superior
the correction fusion module (CFM) upon the first coarse position-
performance compared to the simple concatenation method (the
ing stage, i.e., PVT+FPM+High-res Preserving. If we introduce the
third-row result of Table 2). The good performance reflects that pro-
low-level RGB-aware feature with high resolution to guide the
gressively fusing the prior coarse mask and cross-layer interaction
refining process, we can find that the network outperforms the
is beneficial for camouflaged object refining.
one PVT+FPM. The reason why we need a high-resolution pre-
In summary, the frequency-guided coarse positioning stage mainly
serving module for fine localization lies in two aspects, i.e., 1) the
highlights the important regions of the camouflaged objects under
scales of camouflaged objects are various, and 2) the boundaries
the guidance of hierarchy frequency-aware semantic information,
of camouflaged objects are usually meticulous which are hard to
and the detail-preserving fine localization stage further assists in
discern through the high-level semantic features. Inspired by the
separating the camouflaged objects from the obscure boundaries of
human visual perception system, humans usually need to zoom
the complex background by integrating the high-resolution clue, ad-
in on subtle details in a clear, high-resolution view to recognize
jacent correlation features, and the coarse prediction mask. Finally,
camouflaged objects. If the scale is small in the image, we need to
the proposed FPNet leads us to an accurate and effective solution
leverage the low-level edge-aware or shape-aware information to
for detecting camouflaged objects.
help the network obtain a fine localization. For the obscure bound-
ary problem, multi-scale features fused in a step-by-step manner 4.3.2 Detail Analysis of the Frequency-aware Information.
will give more help to the boundary separating from the complex In order to verify the effectiveness of the frequency perception
background. Thus, we design the refining mechanism to integrate mechanism, we analyze different frequency fusing types through
the high-resolution information and gradually fuse deep features quantitative results, as shown in Table 3. We also provide some
together to solve these problems. The experimental results also visualization comparison results from the prediction mask and the
show that the high-resolution preserving part can provide more learned frequency features, as shown in Figures 6 and 7.
performance gains for detail refining. And we can conclude that the The proposed frequency perception module can automatically
detail refinement strategy is not only significant but also effective separate the features into high-frequency and low-frequency re-
in localizing the camouflaged objects. lated features. However, how to choose a suitable way to integrate
MM ’23, October 29-November 3, 2023, Ottawa, ON, Canada Runmin Cong et al.

Table 3: Quantitative results of frequency-aware features on

the COD10k-Test dataset.

COD10k-Test (2026 images)

Ver. Method
𝑆𝛼 ↑ 𝐸𝑚𝑒𝑎𝑛 ↑ 𝐹 𝛽𝜔 ↑ 𝑀 ↓
No.1 Low-frequency 0.765 0.831 0.438 0.061
No.2 High-frequency 0.848 0.910 0.747 0.029
No.3 High-fre+Low-fre 0.850 0.913 0.748 0.029

the prominent frequency-aware features to help obtain satisfactory Figure 7: Visualization of the learned features about the high-
camouflaged object masks needs further discussion. To verify it, frequency and low-frequency groups. (a) Input image. (b)
we design different comparisons, i.e., only using high-frequency GT. (c) Low-frequency features. (d) High-frequency features.
or low-frequency branches for following camouflaged object mask (e) Features fusing high-frequency and low-frequency after
prediction. The detailed comparisons on the COD10k-Test dataset octave convolution.
are shown in Table 3. We can observe that the only high-frequency
method gives more help than the low-frequency method. All met-
rics scores of only high-frequency outperform the low-frequency the low-frequency and high-frequency groups (Figure 7(e)), our
to a great extent. This reflects that high-frequency information model can focus on the crucial regions of the camouflaged object
is more robust and distinguishable for recognizing camouflaged despite it is similar to the surrounding region.
objects. It also meets the human visual system, we usually em- In conclusion, the proposed frequency perception network has
ploy high-frequency clues to discern the target object from the been verified by analyzing the qualitative and quantitative compar-
uncertain region. However, since the octave convolution is an unsu- ison results that the frequency information can give more help to
pervised operation, that is, no frequency-labeled maps by humans camouflaged object detection. And the proposed frequency percep-
are used for optimization. Thus, some features learned from the tion module can be plugged and played into arbitrary frameworks.
low-frequency branch may be useful for camouflaged object detec-
tion. Moreover, we adopt a simple addition operation fusing the 5 CONCLUSION
high-frequency and low-frequency together, the result is shown in In this paper, we propose a frequency-perception network (FP-
the last row of Table 3. The simple addition of high and low fre- Net) to address the challenge of camouflaged object detection by
quencies achieves the best performance over only single-frequency incorporating both RGB and frequency domains. Specifically, a
ones. Based on these observations, we suggest combining the high- frequency-perception module is proposed to automatically sepa-
frequency and low-frequency into a single addition to obtain further rate frequency information leading the model to a good coarse
improvements. mask at the first stage. Then, a detail-preserving fine localization
In Figure 6, we analyze the influence of different frequency- module equipped with a correction fusion module is explored to
aware feature types. In particular, the prediction masks of only refine the coarse prediction map. Comprehensive comparisons and
high-frequency, only low-frequency, and ours (high-frequency+low- ablation studies on three benchmark COD datasets have validated
frequency) are shown. The high-frequency method can predict the the effectiveness of the proposed FPNet. This work will benefit
key part of the camouflaged object, the low-frequency method can more sophisticated algorithms exploiting frequency clues pursuing
obtain an intact region but with some interference background appropriate solutions in various areas of the multimedia community.
regions. Our proposed method can obtain an accurate object mask In addition, the long-tail problem also exists in COD, this motivates
compared with the high-frequency or low-frequency ones. The com- us to explore reasonable solutions referring to the typical methods
parison results indicate that the frequency features are meaningful of long-tail recognition [59, 60].
for camouflaged object detection. And fusing the high-frequency
and low-frequency will further assist the model in obtaining a ACKNOWLEDGMENTS
relatively complete object mask. This work was supported in part by the National Key R&D Pro-
We also visualize the learned frequency-aware features via the gram of China under Grant 2021ZD0112100, in part by the National
octave convolution to further explain the effectiveness of the pro- Natural Science Foundation of China under Grant 62002014, Grant
posed frequency perception mechanism, as shown in Figure 7. First, 62202461, Grant 62271180, Grant U1913204, Grant U1936212, Grant
our proposed frequency perception mechanism can automatically 62120106009, in part by the Taishan Scholar Project of Shandong
separate the frequency features into high and low frequency groups Province under Grant tsqn202306079, in part by Young Elite Sci-
without any frequency supervision information. Second, we can entist Sponsorship Program by the China Association for Science
observe that the high-frequency and low-frequency groups in the and Technology under Grant 2020QNRC001, in part by the Project
learning process of octave convolution extract the edge information for Self-Developed Innovation Team of Jinan City under Grant
and the main part of the image, respectively. The low-frequency 2021GXRC038, in part by CAAI-Huawei MindSpore Open Fund,
group (Figure 7(c)) focuses more on the overall composition of the and in part by China Postdoctoral Science Foundation under Grant
image, while the high-frequency group (Figure 7(d)) portrays the 2022M723364.
edge part of the camouflaged object in the image. While combing
Frequency Perception Network for Camouflaged Object Detection MM ’23, October 29-November 3, 2023, Ottawa, ON, Canada

REFERENCES 2626–2637.
[1] Nagappa U Bhajantri and P Nagabhushan. 2006. Camouflage defect identification: [23] Ranran Feng and Balakrishnan Prabhakaran. 2013. Facilitating fashion camou-
A novel approach. In 9th International Conference on Information Technology (ICIT). flage art. In ACM International Conference on Multimedia (ACM MM). 793–802.
145–148. [24] Lionel Gueguen, Alex Sergeev, Ben Kadlec, Rosanne Liu, and Jason Yosinski.
[2] Geng Chen, Si-Jie Liu, Yu-Jia Sun, Ge-Peng Ji, Ya-Feng Wu, and Tao Zhou. 2022. 2018. Faster neural networks straight from JPEG. Advances in Neural Information
Camouflaged object detection via context-aware cross-level fusion. IEEE Trans- Processing Systems (NIPS) 31 (2018).
actions on Circuits and Systems for Video Technology 32, 10 (2022), 6981–6993. [25] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross B. Girshick. 2017. Mask
[3] Rui Chen, Fei Li, Ying Tong, Minghu Wu, and Yang Jiao. 2023. A weighted block R-CNN. IEEE Transactions on Pattern Analysis and Machine Intelligence 42 (2017),
cooperative sparse representation algorithm based on visual saliency dictionary. 386–397.
CAAI Transactions on Intelligence Technology 8, 1 (2023), 235–246. [26] Jianqin Yin Yanbin Han Wendi Hou and Jinping Li. 2011. Detection of the mobile
[4] Yunpeng Chen, Haoqi Fan, Bing Xu, Zhicheng Yan, Yannis Kalantidis, Marcus object with camouflage color under dynamic background based on optical flow.
Rohrbach, Shuicheng Yan, and Jiashi Feng. 2019. Drop an octave: Reducing Procedia Engineering 15 (2011), 2201–2205.
spatial redundancy in convolutional neural networks with octave convolution. [27] Junkang Hu, Qiuping Jiang, Runmin Cong, Wei Gao, and Feng Shao. 2021. Two-
In IEEE/CVF International Conference on Computer Vision (ICCV). 3435–3444. Branch Deep Neural Network for Underwater Image Enhancement in HSV Color
[5] Zuyao Chen, Qianqian Xu, Runmin Cong, and Qingming Huang. 2020. Global Space. IEEE Signal Process. Lett. 28 (2021), 2152–2156.
context-aware progressive aggregation network for salient object detection. In [28] Yawen Huang, Feng Zheng, Runmin Cong, Weilin Huang, Matthew R. Scott, and
Thirty-Fourth AAAI Conference on Artificial Intelligence (AAAI). 10599–10606. Ling Shao. 2020. MCMT-GAN: Multi-Task Coherent Modality Transferable GAN
[6] Runmin Cong, Ke Huang, Jianjun Lei, Yao Zhao, Qingming Huang, and Sam for 3D Brain Image Synthesis. IEEE Trans. Image Process. 29 (2020), 8187–8198.
Kwong. early access, doi: 10.1109/TNNLS.2022.3233883. Multi-projection fusion [29] Iván Huerta, Daniel Rowe, Mikhail Mozerov, and Jordi Gonzàlez. 2007. Improving
and refinement network for salient object detection in 360◦ omnidirectional background subtraction based on a casuistry of colour-motion segmentation
image. IEEE Transactions on Neural Networks and Learning Systems (early access, problems. In Third Iberian Conference on Pattern Recognition and Image Analysis.
doi: 10.1109/TNNLS.2022.3233883). 475–482.
[7] Runmin Cong, Jianjun Lei, Huazhu Fu, Ming-Ming Cheng, Weisi Lin, and Qing- [30] Dong Jing, Shuo Zhang, Runmin Cong, and Youfang Lin. 2021. Occlusion-aware
ming Huang. 2019. Review of visual saliency detection with comprehensive bi-directional guided network for light field salient object detection. In ACM
information. IEEE Transactions on Circuits and Systems for Video Technology 29, International Conference on Multimedia (ACM MM). 1692–1701.
10 (2019), 2941–2959. [31] Ch Kavitha, B Prabhakara Rao, and A Govardhan. 2011. An efficient content
[8] Runmin Cong, Qinwei Lin, Chen Zhang, Chongyi Li, Xiaochun Cao, Qingming based image retrieval using color and texture of image sub blocks. International
Huang, and Yao Zhao. 2022. CIR-Net: Cross-modality interaction and refinement Journal of Engineering Science and Technology 3, 2 (2011), 1060–1068.
for RGB-D salient object detection. IEEE Transactions on Image Processing 31 [32] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. 2017. Imagenet classi-
(2022), 6800–6815. fication with deep convolutional neural networks. Commun. ACM 60, 6 (2017),
[9] Runmin Cong, Qi Qin, Chen Zhang, Qiuping Jiang, Shiqi Wang, Yao Zhao, and 84–90.
Sam Kwong. 2023. A weakly supervised learning framework for salient object [33] Trung-Nghia Le, Tam V Nguyen, Zhongliang Nie, Minh-Triet Tran, and Aki-
detection via hybrid labels. IEEE Transactions on Circuits and Systems for Video hiro Sugimoto. 2019. Anabranch network for camouflaged object segmentation.
Technology 33, 2 (2023), 534–548. Computer Vision and Image Understanding 184 (2019), 45–56.
[10] Runmin Cong, Weiyu Song, Jianjun Lei, Guanghui Yue, Yao Zhao, and Sam Kwong. [34] Chongyi Li, Runmin Cong, Sam Kwong, Junhui Hou, Huazhu Fu, Guopu Zhu,
2023. PSNet: Parallel symmetric network for video salient object detection. IEEE Dingwen Zhang, and Qingming Huang. 2021. ASIF-Net: Attention steered inter-
Trans. Emerg. Topics Comput. Intell. 7, 2 (2023), 402–414. weave fusion network for RGB-D salient object detection. IEEE Transactions on
[11] Runmin Cong, Haowei Yang, Qiuping Jiang, Wei Gao, Hai-Sheng Li, Cong Wang, Cybernetics 50, 1 (2021), 88–100.
Yao Zhao, and Sam Kwong. 2022. BCS-Net: Boundary, context, and semantic [35] Chongyi Li, Runmin Cong, Yongri Piao, Qianqian Xu, and Chen Change Loy. 2020.
for automatic COVID-19 lung infection segmentation from CT images. IEEE RGB-D salient object detection with cross-modality modulation and selection. In
Transactions on Instrumentation and Measurement 71 (2022), 1–11. European Conference on Computer Vision (ECCV). 225–241.
[12] Runmin Cong, Ning Yang, Chongyi Li, Huazhu Fu, Yao Zhao, Qingming Huang, [36] Tsung-Yi Lin, Piotr Dollár, Ross B. Girshick, Kaiming He, Bharath Hariharan, and
and Sam Kwong. 2023. Global-and-local collaborative learning for co-Salient Serge J. Belongie. 2016. Feature pyramid networks for object detection. IEEE
object detection. IEEE Transactions on Cybernetics 53, 3 (2023), 1920–1931. Conference on Computer Vision and Pattern Recognition (CVPR) (2016), 936–944.
[13] Runmin Cong, Kepu Zhang, Chen Zhang, Feng Zheng, Yao Zhao, Qingming [37] Jiawei Liu, Jing Zhang, and Nick Barnes. 2021. Confidence-aware learning for
Huang, and Sam Kwong. early access, doi: 10.1109/TMM.2022.3216476. Does camouflaged object detection. arXiv preprint arXiv:2106.11641 (2021).
Thermal really always matter for RGB-T salient object detection? IEEE Transac- [38] Songtao Liu, Di Huang, et al. 2018. Receptive field block net for accurate and fast
tions on Multimedia (early access, doi: 10.1109/TMM.2022.3216476). object detection. In European Conference on Computer Vision (ECCV). 385–400.
[14] Runmin Cong, Yumo Zhang, Ning Yang, Haisheng Li, Xueqi Zhang, Ruochen Li, [39] Zhou Liu, Kaiqi Huang, and Tieniu Tan. 2012. Foreground object detection using
Zewen Chen, Yao Zhao, and Sam Kwong. 2022. Boundary guided semantic larning top-down information based on EM framework. IEEE Transactions on Image
for real-time COVID-19 lung infection segmentation system. IEEE Transactions Processing 21, 9 (2012), 4204–4217.
on Consumer Electronics 68, 4 (2022), 376–386. [40] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin,
[15] Michelle Dean, Robin Harwood, and Connie Kasari. 2017. The art of camouflage: and Baining Guo. 2021. Swin Transformer: Hierarchical vision transformer using
Gender differences in the social behaviors of girls and boys with autism spectrum shifted windows. In IEEE/CVF International Conference on Computer Vision (ICCV).
disorder. Autism 21, 6 (2017), 678–689. 10012–10022.
[16] Fan Deng-Ping, J Ge-Pen, Cheng Ming-Ming, and Shao Ling. 2022. Concealed [41] Yunqiu Lv, Jing Zhang, Yuchao Dai, Aixuan Li, Nick Barnes, and Deng-Ping Fan.
object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence 2023. Towards Deeper Understanding of Camouflaged Object Detection. IEEE
44, 10 (2022), 6024–6042. Transactions on Circuits and Systems for Video Technology (2023).
[17] Max Ehrlich and Larry S Davis. 2019. Deep residual learning in the JPEG transform [42] Yunqiu Lv, Jing Zhang, Yuchao Dai, Aixuan Li, Bowen Liu, Nick Barnes, and
domain. In IEEE/CVF International Conference on Computer Vision (ICCV). 3484– Deng-Ping Fan. 2021. Simultaneously localize, segment and rank the camouflaged
3493. objects. In IEEE/CVF Conference on Computer Vision and Pattern Recognition
[18] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. 2017. Structure- (CVPR). 11591–11601.
measure: A new way to evaluate foreground maps. In IEEE International Confer- [43] Yuxin Mao, Jing Zhang, Zhexiong Wan, Yuchao Dai, Aixuan Li, Yun-Qiu Lv,
ence on Computer Vision (ICCV). 4548–4557. Xinyu Tian, Deng-Ping Fan, and Nick Barnes. 2021. Transformer transforms
[19] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, Ming-Ming Cheng, and Ali Borji. salient object detection and camouflaged object detection. ArXiv abs/2104.10127
2018. Enhanced-alignment measure for binary foreground map evaluation. arXiv (2021).
preprint arXiv:1805.10421 (2018). [44] Ran Margolin, Lihi Zelnik-Manor, and Ayellet Tal. 2014. How to evaluate fore-
[20] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and ground maps?. In IEEE Conference on Computer Vision and Pattern Recognition
Ling Shao. 2020. Camouflaged object detection. In IEEE/CVF Conference on (CVPR). 248–255.
Computer Vision and Pattern Recognition (CVPR). 2777–2787. [45] Haiyang Mei, Ge-Peng Ji, Ziqi Wei, Xin Yang, Xiaopeng Wei, and Deng-Ping Fan.
[21] Deng-Ping Fan, Ge-Peng Ji, Tao Zhou, Geng Chen, Huazhu Fu, Jianbing Shen, and 2021. Camouflaged object segmentation with distraction mining. In IEEE/CVF
Ling Shao. 2020. PraNet: Parallel reverse attention network for polyp segmenta- Conference on Computer Vision and Pattern Recognition (CVPR). 8772–8781.
tion. In Medical Image Computing and Computer Assisted Intervention (MICCAI). [46] Min Ni, Jianjun Lei, Runmin Cong, Kaifu Zheng, Bo Peng, and Xiaoting Fan.
263–273. 2017. Color-Guided Depth Map Super Resolution Using Convolutional Neural
[22] Deng-Ping Fan, Tao Zhou, Ge-Peng Ji, Yi Zhou, Geng Chen, Huazhu Fu, Jianbing Network. IEEE Access 5 (2017), 26666–26672.
Shen, and Ling Shao. 2020. Inf-Net: Automatic COVID-19 lung infection seg- [47] Andrew Owens, Connelly Barnes, Alex Flint, Hanumant Singh, and William
mentation from CT images. IEEE Transactions on Medical Imaging 39, 8 (2020), Freeman. 2014. Camouflaging an object from many viewpoints. In IEEE Conference
on Computer Vision and Pattern Recognition (CVPR). 2782–2789.
MM ’23, October 29-November 3, 2023, Ottawa, ON, Canada Runmin Cong et al.

[48] Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. 2022. A APPENDIX
Zoom in and out: A mixed-scale triplet network for camouflaged object detection.
In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2160– A.1 Qualitative Comparison with SOTA
2170.
[49] Przemysław Skurowski, Hassan Abdulameer, J Błaszczyk, Tomasz Depta, Adam Methods
Kornacki, and P Kozieł. 2018. Animal camouflage analysis: Chameleon database. In Figure 8, we provide more visual examples of different methods.
(2018). [Link]
analysis/ We can see that our proposed network is still competitive in chal-
[50] Yujia Sun, Geng Chen, Tao Zhou, Yi Zhang, and Nian Liu. 2021. Context-aware lenging and difficult scenarios, such as multiple objects, fine objects
cross-level fusion network for camouflaged object detection. In Thirtieth Interna-
tional Joint Conference on Artificial Intelligence (IJCAI). 1025–1031.
and complex background distractions.
[51] Yujia Sun, Shuo Wang, Chenglizhao Chen, and Tian-Zhu Xiang. 2022. Boundary- When there are multiple camouflaged objects, as shown in the
guided camouflaged object detection. arXiv preprint arXiv:2207.00794 (2022). last two rows of Figure 8, other SOTA methods either cannot iden-
[52] Ariel Tankus and Yehezkel Yeshurun. 2001. Convexity-based visual camouflage
breaking. Computer Vision and Image Understanding 82, 3 (2001), 208–237.
tify all the camouflaged objects well or the boundaary of the recog-
[53] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, nized object are not clear. In contrast, our proposed FPNet network
Tong Lu, Ping Luo, and Ling Shao. 2021. Pyramid vision transformer: A versatile can predict all the camouflaged objects with clear and sharp bound-
backbone for dense prediction without convolutions. In IEEE/CVF International
Conference on Computer Vision (ICCV). 568–578. ary.
[54] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong The detection results of our model have a clear advantage in
Lu, Ping Luo, and Ling Shao. 2022. PVT v2: Improved baselines with Pyramid
Vision Transformer. Computational Visual Media 8, 3 (2022), 415–424.
describing the fine details of camouflaged objects. For example, the
[55] Jun Wei, Shuhui Wang, and Qingming Huang. 2020. F3 Net: Fusion, feedback and camouflaged object in the fifth image of Figure 8 has many small,
focus for salient object detection. In AAAI conference on artificial intelligence, thin, burr-like structures. Compared with other methods, only our
Vol. 34. 12321–12328.
[56] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. 2018. CBAM:
method not only accurately detects the camouflaged object, but
Convolutional block attention module. In European Conference on Computer also fully characterizes these trivial details. Similar advantages
Vision (ECCV). 3–19. are also reflected in the situation that the camouflaged target is
[57] Zhe Wu, Li Su, and Qingming Huang. 2019. Cascaded partial decoder for fast
and accurate salient object detection. In IEEE/CVF Conference on Computer Vision occluded. For example, in the images seventh, eighth, eleventh rows
and Pattern Recognition (CVPR). 3907–3916. of Figure 8, the camouflaged object is partially occluded, but our
[58] Fan Yang, Qiang Zhai, Xin Li, Rui Huang, Ao Luo, Hong Cheng, and Deng-Ping method is still robust to this case. It is worth mentioning that we
Fan. 2021. Uncertainty-guided transformer reasoning for camouflaged object
detection. In IEEE/CVF International Conference on Computer Vision (ICCV). 4146– also accurately exclude non-camouflaged occluded object regions
4155. from the prediction results.
[59] Zhiyong Yang, Qianqian Xu, Shilong Bao, Xiaochun Cao, and Qingming Huang.
2022. Learning with multiclass AUC: Theory and algorithms. IEEE Transactions
Furthermore, our method is also able to handle challenging com-
on Pattern Analysis and Machine Intelligence 44, 11 (2022), 7747–7763. plex background scenes, such as the first, second, sixth, ninth and
[60] Zhiyong Yang, Qianqian Xu, Shilong Bao, Yuan He, Xiaochun Cao, and Qingming tenth rows of Figure 8. Taking the sixth row image as an example,
Huang. 2023. Optimizing two-way partial AUC with an end-to-end framework.
IEEE Transactions on Pattern Analysis and Machine Intelligence 45, 8 (2023), 10228– we should detect a ghost pipefish from the image. Other methods
10246. treat the indistinguishable shadows of the ghost pipefish as cam-
[61] Guanghui Yue, Wanwan Han, Bin Jiang, Tianwei Zhou, Runmin Cong, and Tianfu ouflaged object, while only our method can effectively detect the
Wang. 2022. Boundary constraint network with cross layer feature integration
for polyp segmentation. IEEE Journal of Biomedical and Health Informatics 26, 8 ghost pipefish and eliminate interference from shadows.
(2022), 4090–4099.
[62] Chen Zhang, Runmin Cong, Qinwei Lin, Lin Ma, Feng Li, Yao Zhao, and Sam A.2 Visualization of Ablation Studies
Kwong. 2021. Cross-modality discrepant interaction network for RGB-D salient
object detection. In ACM International Conference on Multimedia (ACM MM). We also supplement the visualization results of the ablation ex-
2094–2102.
[63] Qijian Zhang, Runmin Cong, Junhui Hou, Chongyi Li, and Yao Zhao. 2020. CoAD-
periment in Figure 9. Taking the first image as an example, our
Net: Collaborative aggregation-and-distribution networks for co-salient object baseline model (i.e., Figure 9(f)) can only roughly determine the
detection. In Thirty-fourth Conference on Neural Information Processing Systems main part of the camouflaged object, and there is still much room
(NeurIPS). 6959–6970.
[64] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, for improvement in terms of details and accuracy. Further, after the
and Qi Tian. 2023. ControlVideo: Training-free Controllable Text-to-Video Gen- introduction of the frequency-perception module in the baseline
eration. arXiv preprint arXiv:2305.13077 (2023). model, the left shoulder area of the camouflaged human has been
[65] Yijie Zhong, Bo Li, Lv Tang, Senyun Kuang, Shuang Wu, and Shouhong Ding.
2022. Detecting camouflaged object in frequency domain. In IEEE/CVF Conference significantly improved, but the problem of leg integrity remains
on Computer Vision and Pattern Recognition (CVPR). 4504–4513. unresolved. Then, we add a high-resolution preserving design to
our network, which makes the leg details more complete but in-
troduces some noise. Finally, using our designed correction fusion
module in our network allows us to achieve the best performance
with accurate location, complete structure, and sharp boundary.
Frequency Perception Network for Camouflaged Object Detection MM ’23, October 29-November 3, 2023, Ottawa, ON, Canada

Figure 8: Qualitative results of our proposed FPNet model and some state-of-the-art COD methods. The images from left to
right are (a) Input image, (b) GT, (c) Ours, (d) FreNet [65], (e) SINet-V2 [16], (f) PFNet [45], (g) LSR [42], and (h) PraNet [21].

Figure 9: Qualitative results of ablation studies. The images from left to right are (a) Input image, (b) GT, (c) Ours, (d) High-res
Preserving, (e) Frequency-guided Coarse Positioning, (f) Baseline.