Study of an Expansion Method Based on an Image-Specific Classifier and
Multi-Features for Weakly Supervised Semantic Segmentation
Zhengyang Lyu
a
, Pierre Beauseroy
b
and Alexandre Baussard
c
Universit
´
e de Technologie de Troyes, LIST3N, 12 Rue Marie Curie, 10300 Troyes, France
Keywords:
Semantic Segmentation, Weak Supervision, Convolutional Neural Network, Support Vector Machine.
Abstract:
In this paper, we propose a study of an expansion method based on an image-specific classifier and multi-
features for Weakly Supervised Semantic Segmentation (WSSS) with only image-level labels. Recent WSSS
methods focus mainly on enhancing the pseudo masks to improve the segmentation performance by obtain-
ing improved Class Activation Maps (CAM) or by applying post-process methods that combine expansion
and refinement. Most of these methods either lack of consideration for the balance between resolution and
semantics in the used features, or are carried out globally for the whole data set, without taking into account
potential additional improvements based on the specific content of the image. Previously, we proposed an
image-specific expansion method using multi-features to alleviate these limitations. This new study aims
firstly at determining the upper performance limit of the proposed method using the ground truth masks, and
secondly at analysing this performance limit in relation with the features chosen. Experiments show that our
expansion method can achieve promising results, when used with the ground truth (upper performance) and
the features that strike a balance between semantics and resolution.
1 INTRODUCTION
Semantic segmentation is a popular task in computer
vision, with wide applications in various fields such
as autonomous driving, medical imaging, or remote
sensing imaging. However, training a Fully Super-
vised Semantic Segmentation (FSSS) model requires
laborious pixel-level annotations.
Weakly Supervised Semantic Segmentation
(WSSS) has been proposed to reduce the annotation
burden. The weak supervision can be based on
points (Amy et al., 2016), scribbles (Vernaza and
Chandraker, 2017), bounding-boxes (Dai et al.,
2015) or image-level labels (Chen et al., 2022).
The latter, which is considered in this paper, is the
most prevalent in research since it is the easiest and
cheapest annotations to obtain.
Figure 1 illustrates a comparison between the
pipelines of FSSS and WSSS with only image-level
labels. The goal of both tasks is to get a segmenta-
tion model capable of making pixel-level prediction
for a given image. Generally, segmentation mod-
els based on Convolutional Neural Network (CNN)
a
https://orcid.org/0009-0001-8838-4170
b
https://orcid.org/0000-0002-2883-1303
c
https://orcid.org/0000-0002-6693-4282
(Chen et al., 2017; Chen et al., 2018) or more re-
cently, transformers (Strudel et al., 2021) are com-
monly used. In contrast to FSSS, where the ground
truth is available during training, WSSS methods
must first generate pseudo masks, which are used as
ground truth during the second step to train the seg-
mentation model. To generate the pseudo masks, we
start by training a a classification model with image-
level labels and then, generating Class Activation
Maps (CAM) by processing the class-wise deep fea-
tures from the trained network for each image in the
training set (Zhou et al., 2016). Next, an expansion
method, which includes interpolation and argmax op-
erations, is used to get seed. Finally a refinement pro-
cess (Kr
¨
ahenb
¨
uhl and Koltun, 2011; Ahn et al., 2019)
allows to provide an improved pseudo mask, by re-
covering details using characteristics given by image
color features. Of course in WSSS approaches the
quality of the pseudo masks directly influences the ac-
curacy of the segmentation results, since they are used
to train, in the second step, a fully supervised segmen-
tation model. That is why, recent WSSS methods fo-
cus on the generation of pseudo masks whose quality
approaches that of the ground truth.
Figure 2 shows the mean Intersection-over-Union
(mIoU), obtained by several WSSS methods: (Zhang
402
Lyu, Z., Beauseroy, P. and Baussard, A.
Study of an Expansion Method Based on an Image-Specific Classifier and Multi-Features for Weakly Supervised Semantic Segmentation.
DOI: 10.5220/0012462100003654
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 13th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2024), pages 402-409
ISBN: 978-989-758-684-2; ISSN: 2184-4313
Proceedings Copyright © 2024 by SCITEPRESS – Science and Technology Publications, Lda.

Figure 1: Pipelines of fully supervised semantic segmenta-
tion (FSSS) and weakly supervised semantic segmentation
(WSSS) with image-level labels. Lseg is the segmentation
loss function. Lcls is the classification loss function.
Figure 2: mIoU (%) of the pseudo mask on train set (blue
line), mIoU of segmentation (red line), trained from the
pseudo masks, on test set of PASCAL VOC 2012 for several
WSSS methods.
et al., 2020; Li et al., 2022b; Li et al., 2022a; Chen
et al., 2022; Qin et al., 2022; Lee et al., 2022a;
Lee et al., 2021a; Lee et al., 2021b; Jo et al., 2023;
Lee et al., 2022b), for the pseudo masks (blue line)
and the final segmentation (red line). All the con-
sidered methods use a classic expansion process and
common used refinements processes, such as dCRF
(Kr
¨
ahenb
¨
uhl and Koltun, 2011) and IRNet (Ahn et al.,
2019), to generate the pseudo masks. DeepLabV2
(Chen et al., 2017), with ResNet101 backbone, is used
as the fully supervised segmentation model. The qual-
ity of the pseudo-mask is still far from perfect, and
segmentation performance is expected to be better.
Figure 3 provides some visual comparisons be-
tween the outputs of the segmentation models trained
using ground truth in the FSSS task, or the pseudo
mask in the WSSS task. The model trained using
pseudo masks tends to generate predictions with im-
precise and blurred boundaries even if those pseudo
masks seem relatively accurate. However, this effect
is mitigated with full supervision (as can be seen line
3 in figure 3). This suggests that the features learned
by the WSSS model are adversely affected by the la-
belling errors presented in the pseudo-masks. When
pseudo masks are deduced from a global model, the
labelling errors observed on different pseudo masks
can reinforce each other or have common causes
Figure 3: Visual comparison between the the output of the
segmentation models trained by ground truth and pseudo
mask. From top to bottom: image, ground truth, output of
the fully supervised segmentation network, pseudo mask,
output of segmentation network trained with pseudo mask.
and contribute to degrading the overall quality of the
pseudo masks used to create the segmentation net-
work training set (i.e., rail tracks with train). To re-
duce this effect, it is necessary to make better use
of the individual properties of each image and to im-
prove the pseudo masks.
According to those observations, we proposed in
a previous work an image-specific expansion method
using multi-features to alleviate the limitations of ex-
pansion methods which ignore the balance of resolu-
tion and semantics in the features used or miss con-
sideration of image’s specificity. The detailed pro-
posed pipeline is shown in Figure 5. a. We do not
give too specific introduction for this method in this
paper. Just in brief, with only image-level labels, we
designed a sample selection strategy by using multi-
features: CAM, seed and shallow features, to select
data from high-resolution shallow features with suf-
ficient semantics, and label them by the value in the
seed, which enables to create a pixel-wise data set for
training an image-specific Support Vector Machine
(SVM) classifier to infer the pixel-wise prediction for
the entire image. Using this expansion method, we
can get better predictions compared with the original
seed. With further refinement process, our enhanced
pseudo masks are also promising. Some predictions
examples from the SVM are shown in the last column
in Figure 4.
The critical part of the proposed expansion
method is the sample selection process. It directly in-
fluences the accuracy of the labels in the pixel-wise
train set and thus the quality of the prediction. In
order to evaluate the full potential of our expansion
Study of an Expansion Method Based on an Image-Specific Classifier and Multi-Features for Weakly Supervised Semantic Segmentation
403

Figure 4: Qualitative results from the proposed expansion
method in weakly supervised. From the left to right: image,
ground truth, CAM, seed and the result from the SVM in
our expansion method.
method, we conduct, in this study, experiments as-
suming that the ground truth is available, which en-
ables us to define a pixel-wise training set free of
labelling error. This training set can be regarded
as the best result for the sample selection process.
Thus, sample selection step of the original method de-
scribed in Figure 5.a is replaced by a uniform random
sampling from features of the classification network
that are labelled using ground truth, as shown in Fig-
ure 5.b.
For the feature, we explore various options by
choosing activation map values from different layers
in the classification network. Next, for each image
on the training set, a SVM is trained to label its pix-
els. We show that, when ground truth is available, the
prediction results from SVM is particularly promis-
ing under the condition that the features used strike a
balance between resolution and semantics.
The rest of the paper is organized as follows: Sec-
tion 2 presents related work about the expansion pro-
cess in the WSSS framework. Section 3 describes the
proposed expansion method when assuming ground
truth is available. Section 4 provides the experimental
setup and substantial results. We conclude and outline
future research directions in Section 5.
2 RELATED WORK
Generally, weakly supervised semantic methods with
only image-level labels requires a 2-step pipeline:
pseudo mask generation and segmentation model
training. Since the quality of the pseudo masks di-
rectly impacts the performance of the segmentation
model, most methods put main efforts on improving
the accuracy of those pseudo masks. These methods
can be divided into 2 groups. The first one tries to ad-
just classification models to obtain improved CAM,
by strategies likes improving training mechanisms
(Wang et al., 2020) and contrastive learning (Yuan
et al., 2023). The main challenge in these methods
is finding an effective connection between the imple-
mented modifications and the resulting enhancement
of CAMs.
The second one aims at designing better ex-
pansion methods and using refinement processes
to obtain high quality pseudo masks from CAM
(Kr
¨
ahenb
¨
uhl and Koltun, 2011; Ahn and Kwak, 2018;
Ahn et al., 2019; Li et al., 2021; Jo et al., 2023).
For example, PMM (Li et al., 2021) strives to over-
come the partial response problem in CAM by using
a smoothing method to expand the localization area,
and generate class-specific background for each im-
age to independently obtain the pixel-wise label when
obtaining the pseudo mask. In the same way, by gen-
erating class-wise centroids prototype from unsuper-
vised features among the whole dataset, the MARS
method (Jo et al., 2023) is proposed to exclude false
activation made by co-occurrences between the back-
ground elements and associated objects in CAM.
Different with fuzzy localization map provided by
CAM, details are comparatively recovered after the
well-designed expansion method with refinement pro-
cess. We observe that improved expansion methods
can be achieved by using high resolution shallow fea-
tures, such as the color information of the original
image, and relying on the specific semantic informa-
tion for the given image. However, we argue that,
color information may not contain sufficient seman-
tics to well represent the class in the image, which
may lead to incorrect predictions (Kr
¨
ahenb
¨
uhl and
Koltun, 2011; Li et al., 2021). Besides, valuable
image-specific feature may not be effectively utilized
in expansion methods which are implemented glob-
ally (Ahn and Kwak, 2018; Ahn et al., 2019; Jo et al.,
2023).
The proposed expansion method in our previous
work introduces an image-specific pixel-wise classi-
fier with multi-features to solve the limitations in the
recent expansion method. In this paper, we want to
study the potential of this expansion method by as-
suming ground truth is available, and find the opti-
mal options for the feature used by revealing the rele-
vance between performance and resolution-semantics
balance.
3 METHOD
In this section, we describe the details of proposed
expansion method for fully supervised semantic seg-
ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods
404

Figure 5: Detailed pipeline of the proposed expansion method for WSSS (a) and its corresponding fully supervised approach
(b). The solid lines are related to training phase of the SVM classifier, and the dot lines are related to the inference phase.
mentation, as shown in Figure 5.b. We implement
this method assuming the availability of ground truth,
with the goal of assessing its potential upper perfor-
mance limit for the WSSS task and finding the clues
about the feature to use.
The proposed method is divided into 2 steps:
Firstly,for each image we construct a training set com-
posed of a data vector, which can be a set of se-
lected pixels of the image itself or from resized fea-
tures of the classification model (outputted activation
maps of a layer in the trained classification network),
and the corresponding training labels given by the
ground truth. Secondly, we start by training a pixel-
wise classifier, here we consider a SVM model, using
the training set. Then, we output pixel-wise predic-
tions by inferring all the pixels of the original image
or the considered resized features of the classification
model (called the test data in what follows). Finally,
the pixel-wise predictions are reshaped into a two-
dimensional prediction mask.
3.1 Construction of the Training Set
and Test Data
The first part in our method is to construct the pixel-
wise training set, which includes training data and
training labels, and the test data.
The training set is constructed by sampling data
from the feature generated in the classification model
or directly from the original image, and labeling the
data by the ground truth. The details processes are
illustrated as followed: firstly, a classification model
is trained by the given images with image-level la-
bels. Assuming that l classes are labelled in the data
set, we define O = {1, 2,·· · , l} as the set of classes
for image-level labels. The data set is a collection of
image I
I
I ∈ R
W ×H×3
with image-level labels Y , with
Study of an Expansion Method Based on an Image-Specific Classifier and Multi-Features for Weakly Supervised Semantic Segmentation
405

Y ⊂ O. Then, we uniform random sample n training
samples for each class in the image, as shown in Fig-
ure 6. Each selected sample S
S
S
′
′
′
i
is characterized by the
i
th
pixel-wise feature deduced from the original im-
age or outputted activation maps S
S
S of a layer in the
trained classification network after bilinear interpola-
tion to image size. Finally, the pixel-wise training set
is built with the (|Y | + 1)n pairs, (S
S
S
′
′
′
i
,M
M
M
i
), where S
S
S
′
′
′
i
is the pixel-wise feature vector and M
M
M
i
is the ground
truth label in pixel i. |Y | is the number of the object
class in the image.
Notice that, in order to avoid training errors
caused by imbalanced data, the number of samples is
kept the same for each object class and background. It
is essential to stress that the accuracy of SVM results
is related to the total number of samples. When using
a large amount of samples, the results will get better,
at the cost of training time. It is also interesting to set
the total number of samples linked to the resolution
of the feature.
Depending on the layer, the resolution of the cho-
sen feature varied from 1/2 to 1/16 of the original
image. The details resolution for the chosen feature
can be seen in Section 4.1. Normally, the shallow
feature generated from the early layers in the classi-
fication network tends to have shallow-semantic in-
formation with high-resolution. In contrast, the deep
features have high-semantic information with low-
resolution. Different features used in generating lo-
calization maps can bring varied performance. Low-
resolution features are prone to result in smoothed
boundaries and missing details (Zhou et al., 2016),
while insufficient semantics can lead to noisy and in-
accurate predictions (Kr
¨
ahenb
¨
uhl and Koltun, 2011).
Thereby, by comparing the predictions given by the
classifiers trained with different features, the appro-
priate balance between the semantics and resolution
for the ideal prediction can be observed.
To construct the so-called test data, we select all
the pixels of the original image or the feature S
S
S
′
′
′
. The
test data serves as the input for the trained pixel-wise
classifier during the inference stage.
3.2 Training and Infer of the Pixel-Wise
Classifier
As already mentioned and shown in Figure 5, the
pixel-wise classifier, implemented as a Support Vec-
tor Machine (SVM) in our study, conducts a training
phase using the constructed image-specific pixel-wise
training set.
During training, the SVM learns to distinguish be-
tween different classes based on the sampled feature
data. The one-vs-all strategy is employed in multi-
Figure 6: Labeled samples from the ground truth. From left
to the right: image, ground truth, and the selected labeled
samples.
class classification task for SVM, where the classifier
is trained for each class against the others.
Once the SVM is trained, the pixel-wise features
from the test data are inferred to assign class labels.
These predictions are then reconstructed into a seg-
mentation mask.
4 EXPERIMENTS
In this section, we first give the details for experimen-
tal settings like dataset, evaluation metrics and imple-
mentation details. Then, we exhibits our experiment
results quantitatively and qualitatively.
4.1 Experimental Settings
Dataset and Evaluation Metrics. Our preliminary
study is evaluated on PASCAL VOC 2012 dataset
(Everingham et al., 2015), which has 20 foreground
ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods
406

Figure 7: Qualitative segmentation results on PASCAL VOC 2012 train set. From left to right: image, ground truth, prediction
when SVM is trained by using the original image as feature, and features generated from the different layers in the backbone
classification network, namely F
1
to F
5
.
Table 1: Comparison of segmentation mIoU (%) scores using different features on the PASCAL VOC 2012 train set. The best
is highlight by bold.
Feature bkg aero bike bird boat bottle bus car cat chair cow table dog horse motor person plant sheep sofa train tv mean
img 86.98 68.79 35.91 64.40 57.83 58.97 68.77 62.19 77.31 58.65 73.76 69.01 71.25 73.22 57.99 68.02 61.07 70.89 74.65 64.80 60.96 65.97
F
1
98.98 96.70 89.97 96.95 96.39 95.61 96.80 96.74 98.32 95.76 97.77 97.74 97.82 97.75 95.22 96.87 96.17 97.63 97.82 96.70 96.13 96.67
F
2
98.65 92.87 85.61 95.38 93.90 95.34 95.66 95.63 98.16 94.89 96.77 96.76 97.23 96.86 93.93 95.85 95.02 97.22 97.47 95.80 95.19 95.44
F
3
98.58 87.62 84.83 93.02 91.31 96.36 96.74 95.63 97.91 95.57 95.73 97.82 96.44 96.31 93.88 96.27 95.31 96.53 98.29 96.52 97.50 95.15
F
4
96.98 78.48 75.53 86.10 84.00 93.96 94.35 91.75 95.39 91.04 87.35 96.74 90.63 88.29 89.19 93.85 91.03 88.22 96.43 93.71 96.69 90.46
F
5
90.19 55.38 66.41 61.51 60.54 77.25 80.48 75.81 82.78 80.74 72.03 91.36 72.52 71.02 73.94 85.84 81.31 72.71 82.62 73.93 86.00 75.92
classes and 1 background class. The official dataset
is split into train set, validation set and test set, which
contains 1464, 1449 and 1456 images, respectively.
There is an additional annotations provided by Se-
mantic Boundary Dataset (Hariharan et al., 2011)
which augment the train set to 10582 images. The
augmented train set is used to train the classification
network, which is able to generate image-specific fea-
tures for the given image. We evaluate the segmen-
tation results on all 1464 images from the train set.
Mean-intersection-over-union is used to evaluate seg-
mentation results.
Implementation Details. In our experiments, we
use the classification network designed in the weakly
supervised semantic segmentation framework called
Self-supervised Image-specific Prototype Exploration
(SIPE) (Chen et al., 2022) to obtain features for the
given image. Following the set in SIPE, we use
ResNet-50 (He et al., 2016) as the backbone classi-
fication network. The architecture of ResNet-50 is di-
vided into 5 stage, each of which consists in the con-
volutional and pooling layers. We named the output
features from each stage in the backbone classifica-
tion network as F
1
to F
5
, whose spatial size is 1/2,
1/4, 1/8, 1/16 and 1/16 of the original image’s size,
respectively. These features are z-score normalized to
prevent certain parts of features from dominating the
training process due to their larger magnitude. For
each category, including background, in the image,
we set n = 2500 to sample, which corresponds to a
small fraction of image pixels.
4.2 Experiment Results
Figure 7 shows the SVM predictions when using var-
ied features from different stages. We observe that,
when training SVM using the original image as fea-
ture, high-resolution results with relatively detailed
object contours are generated. However, due to the
limited semantics there was a higher occurrence of
Study of an Expansion Method Based on an Image-Specific Classifier and Multi-Features for Weakly Supervised Semantic Segmentation
407

misclassifications in background regions, leading to
a relatively imprecise prediction. Using F
5
feature,
which is in low-resolution with richer semantics, the
predicted segmentation boundaries is blurred and less
accurate. Both of these two features have advantages
in terms of resolution or semantics for SVM results,
but the imbalance between the two factors leads to
imperfect results. In comparison, as shown from col-
umn 4 to column 6 in Figure 7, when using shallow
features with sufficient semantics, like F
1
, F
2
and F
3
,
most of the wrong predictions in the background are
clearly avoided and the objects boundaries are quite
clear.
Segmentation results using different features for
each class category and the mean performance in
the PASCAL VOC 2012 train set are shown in Ta-
ble 1. Results given by using features which strike
the balance between resolution and semantics, i.e.,
F
1
, F
2
and F
3
, shows best segmentation performance
in all categories. Using F
1
feature makes the best
performance in most categories, which suggests that
when ground truth is available, the pixel-wise clas-
sifier trained using high-resolution with sufficient se-
mantics features is able to generate clear segmenta-
tion boundaries and reasonably accurate segmentation
results. Besides, we also did experiments by sim-
ply concatenating selected features, i.e., concatenat-
ing more than one features along the channel axis.
However, the output results from the classifier trained
by the concatenated features are close to the results
from the classifier trained by the feature which has the
most number of channels among the selected features.
It reveals that, when using more than one feature, it is
important to find an appropriate method to combine
them, especially taking into account the differences
in feature depths.
The experiments results shows the great poten-
tial for our expansion method for a segmentation
task. Since we implement the expansion method with
ground truth in this paper, the results can be regarded
as the upper performance limit of the proposed ap-
proach. It reveals that segmentation results in our
WSSS task, which corresponds to the one initially tar-
geted, should be improved when the labels of the se-
lected samples are sufficiently precise and the feature
used represents a good compromise between resolu-
tion and semantics.
5 CONCLUSIONS
When we investigated methods to improve WSSS per-
formances, we observed that the main critical part
is to get better pseudo masks. Compared with the
ground truth, the pseudo masks are not perfect yet.
We start from a proposed expansion method made to
improve them by training a pixel-wise SVM classifier
for each image in the training set. In this study, we
use ground truth to label samples, which is regarded
as the best sampling selection process, to evaluate the
potential of this expansion method. We found that
promising prediction is generated when the used fea-
ture keeps balance between semantics and resolution.
It shows that our expansion method has some poten-
tial and that high-resolution shallow features with suf-
ficient semantics brings effective gain in generating
high quality pseudo masks.
Beside these considerations, it also appears that
the performance of WSSS segmentation networks
are close to the performance obtained with pseudo
masks, thus improving pseudo masks is valuable
if the segmentation network is able to take advan-
tage of their quality as shown in figure 8. Indeed,
FSSS DeeplabV2, used as backbone in all considered
WSSS methods, reaches almost 79%. The gain we
can expect with this backbone from perfect pseudo
masks is then significant but not that large. Our ex-
periments show that by improving the sample selec-
tion process, we could reach more than 95% mIoU
for pseudo masks which is worth the effort if using
the best FSSS network that reaches 90.6% mIoU. So
finally, we can conclude that the efforts put to im-
prove pseudo masks should be adapted according to
the expected performance of the backbone network in
FSSS.
We must also mention that another line of re-
search is to consider segmentation models that self
correct the pseudo masks during training so that
post-processing could be avoided leading to less de-
manding methods from computational point of view.
Works in that direction are under investigation.
Figure 8: mIoU (%) of the pseudo mask on train set (red
points), mIoU of segmentation (model trained from the
pseudo masks, blue points) on test set of PASCAL VOC
2012 for several WSSS methods, and mIoU of the segmen-
tation for fully supervised models: DeepLabV2 and Seg-
NeXt (Guo et al., 2022) on train set (red crosses).
ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods
408

REFERENCES
Ahn, J., Cho, S., and Kwak, S. (2019). Weakly supervised
learning of instance segmentation with inter-pixel re-
lations. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition (CVPR).
IEEE.
Ahn, J. and Kwak, S. (2018). Learning pixel-level semantic
affinity with image-level supervision for weakly su-
pervised semantic segmentation. In Proceedings of
the IEEE Conference on Computer Vision and Pattern
Recognition (CVPR).
Amy, B., Olga, R., Vittorio, F., and Li, F. (2016). What’s
the point: Semantic segmentation with point supervi-
sion. In Proceedings of the European Conference on
Computer Vision (ECCV). Springer.
Chen, L., Zhu, Y., Papandreou, G., F. Schroff, F., and Adam,
H. (2018). Encoder-decoder with atrous separable
convolution for semantic image segmentation. In Pro-
ceedings of the European conference on computer vi-
sion (ECCV). Springer.
Chen, L. C., Papandreou, G., Kokkinos, I., Murphy, K., and
Yuille, A. L. (2017). Deeplab: Semantic image seg-
mentation with deep convolutional nets, atrous convo-
lution, and fully connected crfs. IEEE Transactions
on Pattern Analysis and Machine Intelligence.
Chen, Q., Yang, L., Lai, J., and Xie, X. (2022). Self-
supervised image-specific prototype exploration for
weakly supervised semantic segmentation. In Pro-
ceedings of the IEEE Conference on Computer Vision
and Pattern Recognition (CVPR). IEEE.
Dai, J., He, K., and Sun, J. (2015). Boxsup: Exploiting
bounding boxes to supervise convolutional networks
for semantic segmentation. In Proceedings of the In-
ternational Conference on Computer Vision (ICCV).
IEEE.
Guo, M., Lu, C., Hou, Q., Liu, Z., Cheng, M., and Hu,
S. (2022). Segnext: Rethinking convolutional atten-
tion design for semantic segmentation. Proceedings
of the Conference on Neural Information Processing
Systems (NeurIPS), 35.
He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep resid-
ual learning for image recognition. In Proceedings of
the IEEE Conference on Computer Vision and Pattern
Recognition (CVPR). IEEE.
Jo, S., Yu, I., and Kim, K. (2023). Mars: Model-agnostic
biased object removal without additional supervision
for weakly-supervised semantic segmentation. arXiv
preprint arXiv:2304.09913.
Kr
¨
ahenb
¨
uhl, P. and Koltun, V. (2011). Efficient inference
in fully connected crfs with gaussian edge potentials.
In Proceedings of the Conference on Neural Infor-
mation Processing Systems (NeurIPS). Curran Asso-
ciates, Inc.
Lee, J., Choi, J., Mok, J., and Yoon, S. (2021a). Reducing
information bottleneck for weakly supervised seman-
tic segmentation. In Proceedings of the Conference
on Neural Information Processing Systems (NeurIPS).
Curran Associates, Inc.
Lee, J., Kim, E., Mok, J., and Yoon, S. (2022a). Anti-
adversarially manipulated attributions for weakly su-
pervised semantic segmentation and object localiza-
tion. IEEE Transactions on Pattern Analysis and Ma-
chine Intelligence.
Lee, M., Kim, D., and Shim, H. (2022b). Threshold mat-
ters in wsss: Manipulating the activation for the robust
and accurate segmentation model against thresholds.
In Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition (CVPR). IEEE.
Lee, S., Lee, M., Lee, J., and Shim, H. (2021b). Railroad
is not a train: Saliency as pseudo-pixel supervision
for weakly supervised semantic segmentation. In Pro-
ceedings of the IEEE Conference on Computer Vision
and Pattern Recognition (CVPR). IEEE.
Li, J., Jie, Z., Wang, X., Wei, X., and Ma, L. (2022a).
Expansion and shrinkage of localization for weakly-
supervised semantic segmentation. In Proceedings
of the Conference on Neural Information Processing
Systems (NeurIPS). Curran Associates, Inc.
Li, J., Jie, Z., Wang, X., Zhou, Y., Wei, X., and Ma, L.
(2022b). Weakly supervised semantic segmentation
via progressive patch learning. IEEE Transactions on
Multimedia.
Li, Y., Kuang, Z., Liu, L., Chen, Y., and Zhang, W. (2021).
Pseudo-mask matters in weakly-supervised semantic
segmentation. In Proceedings of the International
Conference on Computer Vision (ICCV).
Qin, J., Wu, J., Xiao, X., Li, L., and Wang, X. (2022).
Activation modulation and recalibration scheme for
weakly supervised semantic segmentation. In Pro-
ceedings of the AAAI Conference on Artificial Intel-
ligence (AAAI). MIT Press.
Strudel, R., R.Garcia, , Laptev, I., and Schmid, C. (2021).
Segmenter: Transformer for semantic segmentation.
In Proceedings of the International Conference on
Computer Vision (ICCV). Springer.
Vernaza, P. and Chandraker, M. (2017). Learning random-
walk label propagation for weakly-supervised seman-
tic segmentation. In Proceedings of the IEEE Con-
ference on Computer Vision and Pattern Recognition
(CVPR). IEEE.
Wang, Y., Zhang, J., Kan, M., Shan, S., and Chen, X.
(2020). Self-supervised equivariant attention mech-
anism for weakly supervised semantic segmentation.
In Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition (CVPR).
Yuan, K., SChaefer, G., Lai, Y., Wang, Y., Liu, X., Guan,
L., and Fang, H. (2023). A multi-strategy contrastive
learning framework for weakly supervised semantic
segmentation. Pattern Recognition.
Zhang, D., Zhang, H., Tang, J., Hua, X., and Sun, Q. (2020).
Causal intervention for weakly-supervised semantic
segmentation. In Proceedings of the Conference on
Neural Information Processing Systems (NeurIPS).
Curran Associates, Inc.
Zhou, B., Khosla, A., Lapedriza, A., et al. (2016). Learn-
ing deep features for discriminative localization. In
Proceedings of the International Conference on Com-
puter Vision (ICCV). Springer.
Study of an Expansion Method Based on an Image-Specific Classifier and Multi-Features for Weakly Supervised Semantic Segmentation
409