Benchmarking computer vision models for automated construction waste sorting
Zhiming Dong, Liang Yuan, Bing Yang, Fan Xue, Weisheng Lu*

5

Department of Real Estate and Construction, Faculty of Architecture, The University of Hong
Kong, Pokfulam, Hong Kong
* Corresponding email: wilsonlu@hku.hk
This is the peer-reviewed post-print version of the paper:
Dong, Z., Yuan, L., Yang, B., Xue, F., & Lu, W. (2025). Benchmarking computer vision
models for automated construction waste sorting. Resources, Conservation and Recycling.
Doi: 10.1016/j.resconrec.2024.108026.
The final version of this paper is available at: https://doi.org/10.1016/j.resconrec.2024.108026.
The use of this file must follow the Creative Commons Attribution Non-Commercial No
Derivatives License, as required by Elsevier's policy.

Abstract

10

15

20

Waste sorting is a critical process in construction waste management system. Computer vision
(CV) offers waste sorting automation potential by recognizing waste composition and
instructing robots or other mechanical devices accordingly. However, how the plethora of CV
models developed perform relative to each other remains underexplored, making model
selection challenging for researchers and practitioners. This study aims to benchmark existing
CV models towards automated construction waste segregation. Seventeen models were
selected and trained with unified configuration, and then their performance was evaluated on
the aspect of accuracy, efficiency, and robustness, respectively. In experimental results, BEiT
attained top accuracy (58.31% MIoU) while FastFCN had the best efficiency (12.87ms). SAN
displayed the least standard deviation (4.41%) for robustness evaluation. This research
contributes a reliable reference for CV model selection, advancing automated construction
waste sorting research and practices, and ultimately promoting efficient recycling while
reducing the environmental impact of construction and demolition waste.
Keywords: Construction Waste Management; Waste Sorting; Computer Vision;
Benchmarking; Composition Recognition

25

1

1. Introduction

30

35

40

45

50

55

60

65

70

75

Construction waste, also known as construction and demolition (C&D) waste, is the solid
waste stream arising from construction activities such as new building construction,
renovation, demolition, foundation excavation, and site formation (Lu et al., 2021). Mainly
comprising concrete, bricks, tiles, soil, rock, metals, timber, gypsum, asphalt, plastics, textiles,
and vegetation, C&D waste accounts for a large share of total solid waste generated around
the world (Yuan et al., 2024). For instance, it contributed 37.1% of solid waste across all
sectors in the European Union in 2020 (European Commission, 2023), 67.3% of solid waste
in the United States in 2018 (USEPA, 2023), 44.0% of solid waste in Australia in 2020
(Gov.AU, 2021), and 40.0% of solid waste in China in 2017 (Zhao et al., 2023). Most
components of C&D waste are chemically stable, making it less hazardous than other solid
waste streams (e.g., municipal solid waste and household waste); however, the massive
amount of C&D waste generated from construction activities amplifies impacts on
environment and society, including landfill space occupation (Yuan et al., 2021), embodied
carbon emission (Liu et al., 2023), illegal dumping (Lu, 2019), and so on. Therefore,
construction waste management (CWM) is a growing environmental protection concern
globally.
Recycling is a prioritized CWM strategy as most C&D waste components are recyclable.
Specific recycling methods vary depending on material composition, while construction waste
is usually a mixture of materials when generated at source (Yuan et al., 2021). Waste sorting,
which means separating mixed waste into different composition groups, is therefore a critical
procedure in CWM (Hoornweg & Bhada-Tata, 2012). In practice, contractors either employ
workers to manually sort construction waste on site, or they transport it to off-site sorting
facilities where, on payment of a levy, the bulk waste is separated using specialized
equipment and then the residual mixtures are sorted manually. Manual construction waste
sorting, albeit common, has been widely criticized for its low efficiency and the serious health
risks it poses to workers (Lu & Chen, 2022).
In promoting a shift from manual sorting to automated separation, computer vision (CV)
enables machinery (e.g., robots) to automatically recognize and separate different C&D waste
materials (Lu & Chen, 2022). Scholars have developed diverse CV models for automated
construction waste sorting. These have achieved promising performance by narrowing
specific gaps, such as integrating advanced algorithms to improve recognition accuracy
(Davis et al., 2021; Zhou et al., 2022), infer the overall waste content according to surface
composition identification (Chen et al., 2021), precisely distinguish distinct waste materials’
mixing boundaries (Dong et al., 2022), recognize compositions in complex real-life scenarios
(Lu, Chen, et al., 2022; Yong et al., 2023), reach acceptable performance on unbalanced and
insufficient data (Na et al., 2022), and optimize feature selection to balance recognition
accuracy and speed (Nežerka et al., 2024), making significant contributions to CV-enabled
construction waste sorting.
A large number of CV models have been proposed, yet how these models perform in
comparison to one another remains underexplored. While some model development studies
have conducted cross-comparisons to highlight the strengths of their proposed models, these
are primarily with limited comparison targets (Lu, Chen, et al., 2022; Nežerka et al., 2024) on
task-specific datasets (Demetriou et al., 2023; Dong et al., 2022). The lack of systematic and
unified comparisons between existing models creates a significant research gap.
Academically, this gap hinders researchers from identifying the advantages and drawbacks of
2

80

85

90

95

100

105

110

115

120

125

various CV methods for CWM, slowing progress in the development of more robust and
accurate CV methods for automated construction waste sorting. Practically, this absence of
systematic comparison complicates the selection process for system developers and
practitioners, potentially leading to suboptimal performance in automated waste sorting
systems, which in turn exacerbates waste management issues and increases environmental
impact.
Based on the research gap, the primary research question guiding this study is: How do
existing CV models for construction waste sorting perform against each other in real-life
CWM scenarios. Therefore, the research objective is to systematically benchmark existing CV
models developed for automating construction waste sorting across various measurement
dimensions under a unified standard. Such benchmarking benefiting from a large image
dataset that presents the visual features of construction waste bulks in real-life scenarios, the
result is compared across various measurement dimensions, including accuracy, efficiency,
and robustness, to provide model selection suggestions. Furthermore, additional valuable
opinions are revealed and discussed, such as challenges ahead and future improvement
opportunities. The remainder of the paper is organized as follows. Section 2 reviews previous
related studies, Section 3 introduces the research methodology and Section 4 describes the
benchmarking experiment procedure and result. Section 5 discusses the research contributions
and shortcomings, and conclusions are drawn in Section 6.
2. Literature review
2.1 Computer vision models proposed for construction waste sorting
Waste sorting plays an essential role in CWM, separating the mixed waste by individual
composition and allowing the implementation of appropriate recycling and reuse strategies
(Hoornweg & Bhada-Tata, 2012). Still common today, manual sorting has been criticized for
its low efficiency and the health hazards it presents to workers (Lu & Chen, 2022). To address
these issues, scholars have proposed various strategies to automate the sorting of C&D waste.
Of these technologies, one of the most promising is the implementation of CV models to help
robotic arms detect and separate different construction and demolition waste materials (Lu &
Chen, 2022).
CV models exhibiting a wide range of techniques, architectures, and applications have been
developed for automated construction waste sorting. Some studies focus on classification
tasks using deep convolutional neural networks (Davis et al., 2021) or hybrid approaches
combining visual and physical features (Chen et al., 2021). Others tackle object detection
challenges with improved YOLOv5 models (Zhou et al., 2022) or real-time detection using
single-stage and two-stage detectors (Demetriou et al., 2023). Semantic segmentation is
another prominent approach, with models like the boundary-aware transformer (Dong et al.,
2022) and DeepLabv3 (Lu, Chen, et al., 2022; Yong et al., 2023) being applied to recognize
waste compositions in complex real-world scenarios. Additionally, some studies address
specific challenges, such as handling unbalanced and insufficient data (Na et al., 2022),
comparing feature extraction methods (Nežerka et al., 2024), or calculating waste volume
(Jiang et al., 2022).
2.2 Performance comparison of CV models for waste sorting
Given the broad landscape of CV models and their implementation in C&D waste sorting, it is
crucial to evaluate and compare their performance so that researchers and practitioners in the
construction industry can identify the most effective approaches as a guide for their future
efforts. The studies we reviewed employ a variety of evaluation metrics, which can be
3

130

135

140

145

150

155

160

165

170

categorized into three main groups: accuracy metrics, efficiency metrics, and robustness
metrics. Accuracy metrics, such as overall accuracy and intersection over union (IoU),
measure the correctness of the model’s predictions compared to the ground truth. For instance,
Davis et al. (2021) and Chen et al. (2021) both reported an accuracy of 94% for their
respective models. Yong et al. (2023) achieved an accuracy of 96.3% and an IoU of 74.6%.
Efficiency metrics, such as computing time and frames per second (FPS), evaluate the
computational efficiency of the model, which is critical for real-time applications. Demetriou
et al. (2023) reported an inference speed of less than 30ms for their model. Robustness
metrics, including deviation and standard deviation, assess the model’s ability to correctly
identify positive instances on data with diverse distributions (Dong et al., 2020).
Although existing methods have generally employed multi-dimensional metrics to evaluate
performance, the scope of such comparisons remains limited, and the standards are
inconsistent. Zhou et al. (2022) proposed a novel computer vision method for construction
waste object detection, comparing it with four baseline methods on a self-collected dataset.
The results showed that their method achieved a mean Average Precision of 0.9480.
Similarly, Dong et al. (2022) introduced a construction waste composition recognition method
based on a vision transformer, comparing its performance with four baseline methods on a
dataset collected from waste disposal facilities in Hong Kong, with an improvement of 5.48%
over the baseline methods. Davis et al. (2021) developed a construction waste material
classification method, evaluating its performance by comparing three different
hyperparameter configurations on a self-collected dataset, which yielded 94% accuracy. In
summary, the primary focus of current computer vision research in automated construction
waste sorting is on model development, with evaluations aimed at emphasizing the strengths
of the proposed models (Demetriou et al., 2023). Consequently, variations in datasets,
experimental setups, and evaluation metrics persist (Nežerka et al., 2024), leading to a
significant research gap due to the absence of systematic and unified comparisons among
existing models.
The innovation of this research lies in conducting a systematic and comprehensive
benchmarking study to address the aforementioned limitations and enable meaningful
comparisons. This study utilized both a consistent, representative dataset and standardized
evaluation metrics for unified comparison. It tested a wide range of C&D waste sorting
models to provide a fair and reliable assessment of their performance. By establishing a
common ground for comparison, this benchmarking study offers valuable insights into the
strengths and limitations of different approaches, identifies areas for improvement, and guides
future research efforts in the field of CV-based construction waste sorting.
3. Methodology
In this section, the methodology of conducting the benchmark study will be illustrated. The
overall technical roadmap includes four steps.
1) Dataset preparation. A dataset includes totally 4,396 common local construction
waste images is prepared for evaluation.
2) Model selection. Seventeen models with different popular architecture or technique are
selected.
3) Model training. All models are trained to convergence with unified standard and same
configuration.
4) Performance evaluation. The performance of each model is evaluated on the aspect of
accuracy, efficiency, and robustness.

175

4

180

185

190

195

3.1 Dataset preparation
With the assistance of staff from the Hong Kong Environmental Protection Department
(HKEPD), this research gathered a comprehensive dataset of construction waste truckloads in
Hong Kong (Chen et al., 2022). The raw data is sourced from sensing systems installed at
multiple government disposal facilities, which record different information including
overhead images taken by cameras (DS-2CD2025FWD-IHONG KONG 4 mm from
Hikvision Digital Technology) positioned at the facility's entrance toll gates, along with
detailed attributes of truck-loaded waste, such as gross vehicle weight, net waste weight, and
waste loading height, recorded by additional sensors. All images captured in October 2019
(5,366 images in total) were collected for this research.
Within Hong Kong’s CWM system, materials are predominantly classified into one of two
categories: inert or non-inert (HKEPD, 2011). Inert materials encompass construction debris,
rubble, earth, bitumen, and concrete, while non-inert materials typically consist of bamboo,
timber, vegetation, packaging waste, and other organic substances. To facilitate enhanced
granularity, the dataset incorporates 10 categories derived from common local construction
waste, with seven representing material types and three serving as auxiliary classifications, as
summarized in Table 1. With the taxonomy, all images were annotated with pixel-level
semantic labels through crowdsourcing services. Each annotation was then manually checked.
Finally, by pre-filtering a few low-quality annotations, an annotated dataset of 4,396 images
was constructed. The waste dataset comprises 4,017 images for the training subset, 200
images for the validation subset, and 179 images for the evaluation subset.
Table 1. The waste dataset information categories

Name
Category
Description
Rock
Inert material
Rock/Stone/Rubble/Debris
Gravel
Inert material
Gravel/Concrete/Bricks
Earth
Inert material
Earth/Slurry/Mud
Packaging Non-inert material Packaging/Fabric/Plastic
Wood
Non-inert material
Wood/Cardboard
Others
Non-inert material
Others (non-inert)
Mixed
Mixed material
Mixed
Background
Auxiliary
Background
Grip
Auxiliary
Grip
Truck
Auxiliary
Truck

200

205

210

Palette

3.2 Model selection
As waste objects depicted in images often possess no fixed shape or location, image
classification and object detection methods struggle to obtain sufficiently accurate information
(Dong et al., 2020) for downstream processing of C&D waste sorting. Consequently, semantic
segmentation methods merit exploration for application to C&D waste sorting (Lu & Chen,
2022). For semantic segmentation models, various architectures and techniques to enhance
performance have been extensively explored. In this study, the segmentation method is
broadly divided into three categories based on architecture: convolutional neural network
(CNN)-based, attention-based, and transformer-based (Csurka et al., 2022). Within each
architectural category, researchers have proposed diverse techniques to improve performance.
For each technique in this study, we select one or two representative methods for
benchmarking, enumerated in Table 2.
5

Table 2. Selected models for benchmarking

Index
Name
Technique
1
FCN (Long et al., 2015)
Deconvolution
2
PSPNet (Zhao et al., 2017)
Pyramid pooling
3
DeepLab V3 (Chen et al., 2017)
Atrous spatial pyramid pooling
4
UPerNet (Xiao et al., 2018)
Unified perceptual parsing
5
DeepLab V3+ (Chen et al., 2018)
Atrous spatial pyramid pooling
6
FastFCN (Wu et al., 2019)
Joint pyramid upsampling
7
OCRNet (Yuan et al., 2020)
Object-contextual representation
8
PSANet (Zhao et al., 2018)
Pointwise spatial attention
9
ENCNet (Zhang et al., 2018)
Context encoding module
10
Non-Local Net (Wang et al., 2018)
Non-local operation
11
APCNet (He et al., 2019)
Adaptive context module
12
DANet (Fu et al., 2019)
Dual attention module
13
ISANet (Huang et al., 2019)
Interlaced sparse self-attention
14
ViT (Dosovitskiy et al., 2020)
Transformer
15
BEiT (Bao et al., 2021)
Pretrained transformer
16
SAN (Xu et al., 2023)
Pretrained multimodal
17
Grounded SAM (Ren et al., 2024)
Vision foundational model

215

220

225

230

235

240

Architecture
CNN
CNN
CNN
CNN
CNN
CNN
CNN
Attention
Attention
Attention
Attention
Attention
Attention
Transformer
Transformer
Transformer
Transformer

3.3 Training
We established a unified model training platform for model training. In terms of hardware
configuration, all models were trained and evaluated on a workstation equipped with an i913900KF central processing unit, an RTX 4090 graphics processing unit, and 128GB of
memory. As for the software configuration, the operating system employed is Ubuntu 22.04,
with the Python interpreter version being 3.8. The comprehensive benchmarking experiments
were conducted using MMSegmentation 1.2.2 and PyTorch 2.0, with the entire platform
constructed upon CUDA 11.8.
The training hyperparameters are set to the default configuration of the MMSegmentation
framework, which emphasizes usability, consistency, and standardization, with proven
performance comparable to other codebases (OpenMMLab, 2020). To ensure a fair
comparison, hyperparameters remain consistent across all models. Specifically, we utilize the
stochastic gradient descent optimizer (Amari, 1993) to optimize model parameters. The
PolyLR (Mishra & Sarawadekar, 2019) serves as the learning rate adjustment policy, with
𝑙𝑙𝑙𝑙 = 𝑚𝑚𝑚𝑚𝑚𝑚{𝑙𝑙𝑙𝑙0 × (1 − 𝑖𝑖 ⁄𝑇𝑇𝑖𝑖 )𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝 , 𝑙𝑙𝑙𝑙𝑚𝑚𝑚𝑚𝑚𝑚 }, where the initial learning rate 𝑙𝑙𝑙𝑙0 is set to 10−2, the
polynomial power is set to 0.9, the total number of iterations 𝑇𝑇𝑖𝑖 is set to 8 × 104 , and the
minimum learning rate at the end of scheduling is set to 10−4. Batch size is determined based
on the highest allowable size under the memory constraint. This study consistently uses a
batch size of 4, guided by the memory demands of the BEiT model, to ensure experimental
fairness.
The input image size for segmentation models is set to 512×512, adhering to the default
configuration for model training. Consequently, a transformation is necessary to convert
images from 1920×1080 to 512×512. This process involves randomly resizing the image, then
flipping and cropping it to 512×512 dimensions. Additionally, some preprocessing techniques
are adopted to enhance the model’s generalization ability, including random adjustments to
brightness, contrast, saturation, and color system (OpenMMLab, 2020).

6

245

250

255

260

265

270

275

280

3.4 Evaluation
3.4.1 Accuracy evaluation
To evaluate performance in terms of accuracy, we use the MIoU metric. Widely adopted in
the research community, this comprehensive metric provides an overall measure of the
segmentation accuracy of a model across all classes, making it easier for us to compare and
benchmark different models. MIoU can be computed using Formula (1), where 𝑝𝑝𝑖𝑖𝑖𝑖 denotes
the number of pixels belonging to the 𝑖𝑖𝑡𝑡ℎ category but predicted as the 𝑗𝑗𝑡𝑡ℎ category. 𝑘𝑘
represents the total number of categories.
1

𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀 = 𝑘𝑘 ∑𝑘𝑘𝑖𝑖=1�𝑝𝑝𝑖𝑖𝑖𝑖 ⁄�∑𝑘𝑘𝑗𝑗=1 𝑝𝑝𝑖𝑖𝑖𝑖 + ∑𝑘𝑘𝑗𝑗=1 𝑝𝑝𝑗𝑗𝑗𝑗 − 𝑝𝑝𝑖𝑖𝑖𝑖 ��

(1)

3.4.2 Efficiency evaluation
Computation time is employed to assess efficiency performance. The efficiency of CV
algorithms is a crucial factor to consider in practical applications. Although efficiency may be
less critical in certain situations, such as smartphone-assisted household waste classification,
stringent time performance is necessary in scenarios like automated waste sorting with
robotics. This is particularly relevant when considering that a waste recycling facility can
process thousands of tons of municipal solid waste per day, and sorting speed directly impacts
overall throughput (Lu & Chen, 2022).
3.4.3 Robustness evaluation
The robustness evaluation aims to assess the stability of semantic segmentation methods’
performance on various waste distributions. Lu, Yuan, et al. (2022) found that gross vehicle
weight and vehicle type significantly affect waste distributions. Heavy inert construction
waste is typically transported by dump trucks, while light construction waste is often hauled
by skip container trucks due to their large-sized truck buckets. Skip container trucks can
transport any type of construction waste. Regarding the influence of gross vehicle weight on
waste distribution, heavier trucks are commonly used for transporting heavy construction
waste. Lighter trucks, usually vans, are employed by small renovation companies for
refurbishment or minor works.
This loading pattern can be utilized to divide the evaluation subset into distinct groups.
Vehicle types can be categorized into grab mount, dump, and skip container, while gross
vehicle weight can be split into six types ranging from 9 to 30 tons. Table 3 illustrates the data
split results. There are a total of nine groups; the 16-ton skip container group contains 19
images, and all other groups have 20 images. Consequently, the evaluation subset comprises
179 images. The pie chart in each group displays the data category distribution, revealing
significant differences among categories. During the robustness evaluation, the MIoU of each
group is initially calculated, and the standard deviation across all groups represents the
robustness.

285

7

Table 3. Data category distribution of the evaluation subset
Gross
Proportion of each category (%)
Group Vehicle Vehicle
Index
Type
Weight Rock Gravel Earth Packaging Wood Others Mixed
(ton)
1
Skip
16
4.9
19.1
7.8
13.9
42.3
5.9
6.0
2
container
24
33.8
20.8
3.8
27.5
10.9
0.3
3.0
3
16
11.9
10.0
0.2
7.3
47.1
19.2
4.3
Grab
4
24
5.2
13.5
5.4
20.5
21.9
13.2
20.3
mount
5
30
8.5
3.6
8.8
8.7
33.4
3.0
34.0
6
9
0
1.9
0.4
14.4
61.6
21.6
0.1
7
13
0.5
8.8
0.3
10.0
56.5
16.0
7.9
Dump
8
14
4.7
7.2
1.1
3.6
56.8
22.6
3.9
9
16
2.2
9.8
0
7.6
57.4
21.6
1.4

290

295

Specifically, each model is firstly evaluated by MIoU on each group, and the average of all
those MIoUs is recorded as 𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚 , while the standard deviation is recorded as 𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑆𝑆𝑆𝑆 .
The 𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑆𝑆𝑆𝑆 serves as an indicator of the method’s robustness. The calculation formula is
shown in Formula (2) and (3), where n is the number of loading patterns, n=9 in this case,
𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑖𝑖 refers to the evaluation result on i-th group.
1

𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚 = 𝑛𝑛 ∑𝑛𝑛𝑖𝑖=1 𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑖𝑖

𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑆𝑆𝑆𝑆 = �∑𝑛𝑛𝑖𝑖=1(𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑖𝑖 − 𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚 )2 ⁄𝑛𝑛

300

305

310

(2)
(3)

4. Benchmarking experiments
In this Section, we analyze the benchmarking results of the 17 methods selected for
comparison. It is worth noting that Grounded SAM is used in a zero-shot learning manner,
meaning it is not trained on the training subset and is evaluated with the evaluation subset
directly. All other methods are fully trained on the training subset and then evaluated using
the evaluation subset. The accuracy, efficiency, and robustness performance aspects of all
methods are discussed from Section 4.1 to Section 4.3.
4.1 Accuracy analysis
The overall accuracy evaluation results are presented in Table 4. The header of table contains
the category name and corresponding color palette. The table content displays the evaluation
results. Each row lists the detailed IoU value for each category of a specific method. Among
all 17 methods, BEiT achieved the best performance with an MIoU value of 58.31%. BEiT
performed well in most categories, including background, gravel, packaging, wood, others,
grip, and truck. The vast number of parameters enables BEiT to learn rich, contextual
representations from both local and global image features, therefore contributed to BEiT’s
overall superior results compared to other methods.

315

8

Table 4. Overall accuracy evaluation result.
ID

Background

Rock

Gravel

Earth

Packaging

Wood

Others

Mixed

Grip

Truck

MIoU

FCN

1

93.27

32.04

41.56

35.57

55.57

69.72

39.58

31.89

90.39

77.32

56.691

PSPNet

2

93.09

27.03

39.58

34.29

56.76

69.76

39.18

33.31

90.25

77.5

56.075

DeepLabv3

3

93.28

34.53

41.61

32.81

56.37

70.51

39.28

33.86

90.4

77.35

57

UPerNet

4

93.33

36.13

39.58

36.32

57.11

69.93

38.58

34.21

90.36

76.87

57.242

DeepLabv3+

5

93.34

32.07

44.45

34.57

55.94

69.94

38.61

33.26

90.31

77.48

56.997

FastFCN

6

93.25

33.93

39.25

35.65

56.01

68.34

38.31

33.12

90.12

76.82

56.48

OCRNet

7

93.19

36.27

42.73

38.19

56.12

69.4

39.11

33.28

90.24

76.74

57.527

PSANet

8

93.3

37.85

42.64

36.4

56.47

70.28

40.5

36.34

90.16

77.23

58.117

ENCNet
Non-local
Net
APCNet

9

93.27

35.43

41.72

33.58

55.98

69.68

38.03

35.11

90.24

77.23

57.027

Name
Palette

320

325

330

335

340

345

10

93.28

33.26

42.33

35.91

56

69.82

38.73

35.03

90.32

77.38

57.206

11

93.32

36.36

41.63

33.75

55.35

70

38.57

32.55

90.29

77.24

56.906

DANet

12

93.25

41.2

37.59

39.31

56.41

69.82

37.28

35.49

89.87

76.62

57.684

ISANet

13

93.36

34.23

39.27

31.32

55.85

69.5

40.65

33.57

90.17

77.17

56.509

ViT

14

93.24

34.8

36.26

38.7

55.25

67.69

37.07

33.96

88.33

75.89

56.119

BEiT

15

SAN

16

93.54
92.9

34.02
32.9

42.96
42.1

38.4
36.5

58.26
53.2

71.56
67.8

41.48
36.2

35.05
35.2

90.22
87.3

77.61
72.7

58.31
55.68

Grounded
SAM

17

51.58

5.36

21.07

0.73

2.58

3.85

0.89

0

1.54

14.48

10.208

For the other results, OCRNet achieved the best performance with an MIoU value of 57.53%
when comparing with other models with CNN architecture, such results mainly contributed by
the good performance in the rock and earth categories. PSANet achieved the best performance
across all attention-based methods, with an overall MIoU result of 58.12%. The strong
performances con categories packaging, wood, and mixed contribute positively to PSANet's
overall good performance.
The MIoU of Grounded SAM is 10.21%, a significant performance gap compared with other
fully supervised methods. Despite the overall accuracy performance being subpar, there are
noteworthy aspects that showcase the considerable promise of zero-shot learning methods,
such as Grounded SAM, where the model needs no training phase and can be directly utilized
for C&D waste sorting. For the output of Grounded SAM, some good samples prove it can
pinpoint basic objects and discern their contours. However, failure examples also reveal the
instability of Grounded SAM’s performance, as it struggles with scenarios involving multiple
waste types or catastrophic errors in object category identification.
In real-time waste sorting applications, accuracy is the most crucial factor for successful
operations. Highly accurate models ensure that materials are properly categorized, preventing
contamination and costly errors in waste management. While efficiency and robustness are
important, accuracy remains the cornerstone of effective real-world waste sorting systems.
Engineering managers should prioritize accuracy, even if it requires sacrificing some
efficiency. When trade-offs are necessary between accuracy, efficiency, and robustness, they
must be carefully weighed.
4.2 Efficiency analysis
Computing time is utilized to assess efficiency performance. For the overall evaluation subset,
the average computing time of all images is recorded and employed for evaluation. The
9

evaluation result is depicted in Fig. 1. Fig. 1 is a combination chart includes a bin chart and a
line chart. The MIoU value can be obtained from the primary axis on the left, and it is
displayed in blue bar chart as a percentage. The calculation time value is obtained from the
secondary axis on the right, and it is displayed in orange line chart as milliseconds.
350

Fig. 1. Efficiency evaluation result. The blue bin chart represents the MIoU performance of
each model, while the orange points on the line chart indicate the corresponding computing
time.
355

360

365

370

375

Fig. 1 illustrates that both the CNN-based and attention-based methods consistently achieved
a computing time below 20ms for an input image size of 1920×1080. The FastFCN method
demonstrated the fastest computation time, requiring only 12.87ms to calculate the
segmentation result for a single image. Such fast inference primarily due to its Joint Pyramid
Upsampling module to minimizes the computational complexity and memory usage. Among
all the CNN-based and attention-based methods, OCRNet exhibited the longest computation
time, which amounted to 24.56ms to allow its Object-Contextual Representation module to
capture detailed relationships between pixels and their surrounding context. However, it also
achieved the highest accuracy performance among all the CNN-based and attention-based
methods. In terms of efficiency performance, Grounded SAM’s computing time was
342.88ms, ten times that of CNN-based and attention-based methods due to the computational
complexity of integrating visual segmentation with language grounding. However, such
latency remains acceptable for practical C&D waste sorting tasks (Lu & Chen, 2022).
Efficiency performance holds significant practical importance. Depending on the speed of
sorting hardware, such as conveyors and robotic arms, real-world applications exhibit varying
sensitivities to latency. In high-speed sorting lines where waste moves rapidly, models with
low latency should be prioritized. Lower latency typically allows a CV model to process more
frames per second, accelerating the waste sorting process and enhancing the efficiency and
throughput of the entire system. Conversely, if the system is less sensitive to latency, models
with higher accuracy can be considered to reduce errors in automated waste sorting and
improve overall performance.

10

380

385

390

395

4.3 Robustness analysis
A robust model ensures that the system can reliably handle real-world waste management
tasks, especially C&D waste with the characteristic of bulky and heterogeneous. Therefore, as
detailed in Section 3.4.3, the deviation on different data distribution patterns 𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑆𝑆𝑆𝑆 is used
to evaluate the robustness. There are totally nine group of loading pattern, which is the
combination of different vehicle types and gross vehicle weight, are used to calculate the
deviation.

Fig. 2. Robustness evaluation result. For sub-figure (a), blue bar refers to the MIoU
performance of each model, while the error bar refers to the robustness metric 𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑆𝑆𝑆𝑆
calculated by formula (3). For sub-figure (b), horizonal axis is the accuracy evaluation result
represented by MIoU, vertical axis refers to the robustness evaluation result represented by
𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑆𝑆𝑆𝑆 . The blue dotted line is the trendline of the relationship between MIoU and 𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑆𝑆𝑆𝑆 .

The robustness results are presented in Fig. 2. In Fig. 2 (a), the bars represent the MIoU values
on the total evaluation dataset, while the error bars indicate the corresponding 𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑆𝑆𝑆𝑆 values.
A smaller 𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑆𝑆𝑆𝑆 value indicates better robustness performance. SAN exhibits the smallest
𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑆𝑆𝑆𝑆 of 4.58%, making it the most robust method. ISANet demonstrates the largest
𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑆𝑆𝑆𝑆 value of 10.11%, indicating the lowest level of robustness. Despite BEiT’s
11

400

405

410

415

420

425

430

435

440

445

outstanding accuracy performance, its 𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑆𝑆𝑆𝑆 value stands at 8.83%. Concerning robustness
performance, Grounded SAM’s 𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑆𝑆𝑆𝑆 is 1.49%, the smallest among all methods. However,
such a low 𝑀𝑀𝑀𝑀𝑜𝑜𝑜𝑜𝑆𝑆𝑆𝑆 does not guarantee robustness as its MIoU is also low, indicating that
Grounded SAM’s fluctuation range must remain minimal, resulting in a low 𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑆𝑆𝑆𝑆 value.

Fig. 2(b) illustrates a correlation between robustness and accuracy. Grounded SAM is
excluded from this analysis due to its significantly different accuracy performance. A linear
trendline is applied to fit the correlation, which is calculated using the least squares method.
The results show that MIoU and 𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑆𝑆𝑆𝑆 are positively correlated, indicating that higher
accuracy generally corresponds to lower robustness. This is likely because models highly
optimized for accuracy on a specific dataset tend to overfit to details or noise within that data
(Tsipras et al., 2018). The reliability of this trend is measured by the R-squared value of
0.0774, suggesting the correlation is not absolute, and that the complexity of the model
architecture significantly impacts performance evaluation. There are several noteworthy
deviations from the linear relationship between accuracy and robustness. For example,
OCRNet exhibits the second-best robustness after SAN, with an 𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑆𝑆𝑆𝑆 of 6.7%, yet its
MIoU is 1.85% higher than SAN. PSANet, with the second-highest accuracy at 58.12%,
demonstrates more stable performance in robustness evaluation, with an 𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑆𝑆𝑆𝑆 of 7.9%.

This robustness analysis highlights an important consideration when evaluating accuracy: a
model must not only perform well under optimal conditions but also consistently handle
diverse real-world scenarios. In real-world waste sorting, when waste distribution is relatively
simple and stable over time and space, users can prioritize model accuracy. However, when
waste is highly heterogeneous or bulky, both accuracy and robustness must be considered, and
a balance between the two should be struck. This ensures that the CV model can function
effectively in ideal conditions while providing reliable information for the automatic sorting
of construction waste.
5. Discussion
This research presents a benchmark study that compares the performance of various CV
models in the task of automated construction waste sorting. Compared with model
development studies, this work offers a broader and more objective comparison by evaluating
CV models with different architectures under a unified standard. The key advantage of this
approach is its ability to ensure consistent evaluation across models, enabling more informed
decisions when selecting models for specific applications, thereby enhancing practical
deployment. Compared with general benchmark studies on public datasets, this research
focuses on the specialized task of construction waste sorting, which presents unique
challenges due to domain gaps. Waste objects are highly cluttered and heterogeneous, making
this task significantly different from general vision tasks. The strength of this study lies in
providing benchmark results on a construction waste sorting dataset, offering domain-specific
insights crucial for advancing waste management automation. Based on the comparison
results, valuable insights regarding model selection, upcoming challenges, and future research
recommendations are detailed in Sections 5.1 to 5.3.
5.1 Suggestions for model selection
The evaluation results of all fully supervised methods from this research indicate that in terms
of accuracy, the MIoU varies from 55.68% to 58.31%, with BEiT achieving the best
performance. In terms of efficiency, the computation time ranges from 12.87ms to 203.72ms,
and FastFCN exhibits the best performance. Regarding robustness, the standard deviation of
12

MIoU across different data distribution patterns varies between 4.48% and 10.11%, with SAN
demonstrating the best performance.
450

455

460

465

470

475

480

485

490

495

While the assessment outcome is clear-cut, potential users and researchers should weigh the
tradeoffs among these three aspects for successful implementation. For instance, even though
BEiT boasts the best performance, it requires 204ms to process a single image, over 10 times
longer than that required by CNN-based and attention-based methods, rendering it potentially
unsuitable for time-sensitive C&D waste sorting situations. Consequently, we can offer
recommendations for model selection based on these evaluation findings.
1) General performance: The balanced mode necessitates a compromise between
accuracy, efficiency, and robustness. This widely preferred mode can be considered
the default choice. PSANet is the most fitting model for this mode, as its MIoU is
58.12%, only 0.19% lower than the best performance. It requires 15.28ms for
computation, which is a mere 2.41ms longer than the best performance. The standard
deviation is 7.91%, which is 3.43% higher than the best performance.
2) Accuracy-prioritized: For users who need extremely accurate performance and can
tolerate longer computing times, BEiT is the most suitable model. It has the highest
MIoU of 58.31% and performs well in both overall categories and waste category
evaluations.
3) Efficiency-prioritized: For users who require a CV method for real-time C&D waste
sorting tasks or have limited hardware capabilities, FastFCN meets their needs. It
requires only 12.87ms for computation and achieves a MIoU of 57.53% with a 7.38%
standard deviation.
4) Robustness-prioritized: For users who prioritize prediction stability in C&D waste
sorting tasks with a wide variety of waste types, mixture patterns, or distributions,
SAN is the most suitable option due to its low standard deviation of 4.59%.
5) Annotation-free: For users who lack the resources to annotate custom C&D waste
sorting data or only require basic semantic understanding without any workload,
Grounded SAM is the ideal choice. Based on general visual understanding capabilities,
Grounded SAM can provide semantic segmentation results for C&D waste sorting in a
zero-shot manner.
5.2 Challenges ahead
The CV-based automated waste sorting system can foster a transition from manual sorting to
automated separation. Nonetheless, this benchmarking study identifies certain drawbacks that
could present obstacles in translating current academic endeavors into practical applications.
Firstly, the low quality of C&D waste sorting data hinders the accuracy of CV methods. C&D
waste sorting datasets are often captured under complex real-life conditions where waste
objects lack clear boundaries and fixed shapes. These objects are stacked together, resulting in
construction waste that is bulky, heterogeneous, cluttered, and heavily overlapping (Lu &
Chen, 2022). Even manual identification of construction waste features from images can be
challenging due to these factors (Dong et al., 2022). This presents a challenge for the accuracy
performance of CV methods applied to C&D waste sorting.
The second challenge arises from the inherent data imbalance present in construction waste
data. Due to the varying demands for different types of materials during construction
processes, there are significant disparities in the waste quantity distribution. In cases of
imbalanced data, the model tends to be biased towards the majority class, which has more
13

500

505

510

515

520

525

530

535

540

545

samples to learn from (Dong et al., 2020). This imprecision in C&D waste sorting can result
in inaccurate component estimation. More critically, the majority class often consists of
auxiliary categories such as background, whereas the minority class typically includes waste
types that researchers are more interested in (Lu & Chen, 2022). Although accurately
identifying auxiliary categories can assist the model in locating and recognizing waste, the
data imbalance negatively affects the recognition results in a way that cannot be overlooked.
The final challenge is the time consumption of transformer-based models. These models,
containing billions of parameters, offer benefits such as improved performance and reduced
annotation workload. However, they also present challenges in computational requirements
and memory usage. Evaluation results indicate that CNN and attention-based models typically
take about 15ms for a single image, whereas transformer-based models require around 200ms
for individual tasks. Notably, while experimental results are conducted on an RTX 4090 GPU,
actual computing speeds may vary depending on the hardware, as only a few platforms
possess such powerful GPUs. This discrepancy may result in significant latency for real-world
C&D waste sorting computing tasks.
5.3 Future research recommendations
Considering the current challenges and limitations, future research may explore the following
avenues. First, researchers can address low-quality data through data augmentation, cleaning,
or implementing more robust architectures (Lu & Chen, 2022). Second, data imbalance can be
tackled using resampling techniques, such as over-sampling for minority categories or undersampling for majority categories (Khushi et al., 2021). Researchers are also exploring
modified loss functions to address imbalance, including weighted cross-entropy loss or focal
loss (Dong et al., 2020). Lastly, the time consumption issue of transformer-based vision
foundation models can be managed by deploying more powerful computing platforms or
utilizing cloud services to offload local computing power (Ni et al., 2024). Although
transformer-based methods demonstrate high performance and potential in C&D waste
sorting, users must consider all factors and select the most suitable methods for their specific
tasks. These efforts ultimately contribute to the practical significance of construction waste
management in engineering projects, by reducing environmental impacts, promoting
recycling, curbing illegal dumping, and fostering sustainable economic growth.
6. Conclusions
CV offers significant potential for facilitating the transition from manual sorting to automated
separation. This study aimed to benchmark existing CV models that have been developed for
automating C&D waste sorting, categorizing them into three groups: CNN-based, attentionbased, and transformer-based. We evaluated all models using a unified standard and compare
them in terms of accuracy, efficiency, and robustness. The contributions of this research are:
1) A benchmarking pipeline designed for performance evaluation. The research
methodology for deriving benchmarks consists of dataset preparation, model selection,
training, and evaluation. Each step was carefully designed to establish a universal and
standard pipeline for automated CV-based C&D waste sorting. All hyperparameters
are well controlled for fair comparison. Based on the benchmarking pipeline,
seventeen existing common CV models were evaluated and compared across three
different measurement dimensions, i.e., accuracy, efficiency, and robustness.
2) Upon analysing the evaluation results, this research offers practical suggestions based
on different real-world usage requirements and outlines the challenges faced by
current CV methods for C&D waste sorting, proposing future directions to address
each challenge. The study holds both theoretical significance and practical value.
14

550

555

560

565

570

575

580

585

590

Theoretically, it allows researchers to identify the strengths and weaknesses of various
CV approaches, laying the groundwork for refining existing models and developing
more innovative, robust solutions for waste sorting. From an engineering perspective,
it assists practitioners and system developers in selecting the most suitable CV models
for real-world applications, improving automated sorting performance and enhancing
waste management outcomes. Ultimately, this research fosters more efficient recycling
and reduces the environmental impact of construction and demolition waste.
This pipeline not only ensures that benchmarks are derived in a systematic and rigorous
manner, but also facilitates a fair comparison and understanding of the performance of various
models and techniques within the C&D waste sorting domain.
The benchmarking results reveal that BEiT achieved the highest accuracy performance, with a
MIoU of 58.31%. FastFCN exhibited the shortest computing time, clocking in at 12.87ms
during efficiency evaluation. SAN demonstrated the smallest standard deviation of 4.41% in
terms of robustness evaluation. PSANet emerges as the recommended model due to its
excellent balance across the three aspects, while Grounded SAM offers a straightforward
“load and go” approach to alleviate the annotation workload.
By benchmarking these models against a unified standard, this study identified three key
challenges: low data quality, data imbalance, and the time consumption of transformer-based
models. For each challenge, it proposed future directions to address low-quality data, mitigate
data imbalance, and manage the computational demands of transformer-based models. This
study facilitates informed decision-making by researchers and practitioners in the selection
and development of CV models. Furthermore, it instills confidence in the adoption of CVbased automated C&D waste sorting for normal waste management workflows in the
construction industry.
References
Amari, S.-i. (1993). Backpropagation and stochastic gradient descent method.
Neurocomputing, 5(4-5), 185-196.
Bao, H., Dong, L., Piao, S., & Wei, F. (2021). Beit: Bert pre-training of image transformers.
arXiv preprint arXiv:2106.08254.
Chen, J., Lu, W., & Xue, F. (2021). “Looking beneath the surface”: A visual-physical feature
hybrid approach for unattended gauging of construction waste composition. Journal of
environmental management, 286, 112233.
Chen, J., Lu, W., Yuan, L., Wu, Y., & Xue, F. (2022). Estimating construction waste truck
payload volume using monocular vision. Resources, conservation and recycling, 177,
106013.
Chen, L.-C., Papandreou, G., Schroff, F., & Adam, H. (2017). Rethinking atrous convolution
for semantic image segmentation. arXiv preprint arXiv:1706.05587.
Chen, L.-C., Zhu, Y., Papandreou, G., Schroff, F., & Adam, H. (2018, September 8-14).
Encoder-decoder with atrous separable convolution for semantic image segmentation.
Proceedings of the European conference on computer vision (ECCV), Munich,
Germany.
Csurka, G., Volpi, R., & Chidlovskii, B. (2022). Semantic image segmentation: Two decades
of research. Foundations and Trends® in Computer Graphics and Vision, 14(1-2), 1162.
15

595

600

605

610

615

620

625

630

635

640

Davis, P., Aziz, F., Newaz, M. T., Sher, W., & Simon, L. (2021). The classification of
construction waste material using a deep convolutional neural network. Automation in
construction, 122, 103481.
Demetriou, D., Mavromatidis, P., Robert, P. M., Papadopoulos, H., Petrou, M. F., &
Nicolaides, D. (2023). Real-time construction demolition waste detection using stateof-the-art deep learning methods; single–stage vs two-stage detectors. Waste
Management, 167, 194-203.
Dong, Z., Chen, J., & Lu, W. (2022). Computer vision to recognize construction waste
compositions: A novel boundary-aware transformer (BAT) model. Journal of
Environmental Management, 305, 114405.
Dong, Z., Wang, J., Cui, B., Wang, D., & Wang, X. (2020). Patch-based weakly supervised
semantic segmentation network for crack detection. Construction and Building
Materials, 258, 120291.
Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T.,
Dehghani, M., Minderer, M., Heigold, G., & Gelly, S. (2020). An image is worth
16x16 words: Transformers for image recognition at scale. arXiv preprint
arXiv:2010.11929.
European Commission, E. C. (2023). Database-Waste-Eurostat-European Commission.
Retrieved 1 December 2023 from
https://ec.europa.eu/eurostat/web/waste/data/database
Fu, J., Liu, J., Tian, H., Li, Y., Bao, Y., Fang, Z., & Lu, H. (2019). Dual attention network for
scene segmentation. Proceedings of the IEEE/CVF conference on computer vision and
pattern recognition, Long Beach, CA, United States.
Gov.AU. (2021). National Waste Report 2020. Retrieved 1 December 2023 from
https://shorturl.at/orEKS
He, J., Deng, Z., Zhou, L., Wang, Y., & Qiao, Y. (2019). Adaptive pyramid context network
for semantic segmentation. Proceedings of the IEEE/CVF conference on computer
vision and pattern recognition, Long Beach, CA, United States.
HKEPD. (2011). Construction Waste Statistics. Retrieved 2 from https://www-epd-govhk.eproxy.lib.hku.hk/epd/misc/cdm/trip.htm
Hoornweg, D., & Bhada-Tata, P. (2012). What a waste: a global review of solid waste
management.
Huang, L., Yuan, Y., Guo, J., Zhang, C., Chen, X., & Wang, J. (2019). Interlaced sparse selfattention for semantic segmentation. arXiv 2019. arXiv preprint arXiv:1907.12273.
Jiang, Y., Huang, Y., Liu, J., Li, D., Li, S., Nie, W., & Chung, I.-H. (2022). Automatic
volume calculation and mapping of construction and demolition debris using drones,
deep learning, and GIS. Drones, 6(10), 279.
Khushi, M., Shaukat, K., Alam, T. M., Hameed, I. A., Uddin, S., Luo, S., Yang, X., & Reyes,
M. C. (2021). A comparative performance analysis of data resampling methods on
imbalance medical data. IEEE access, 9, 109960-109975.
Liu, J., Li, Y., & Wang, Z. (2023). The potential for carbon reduction in construction waste
sorting: a dynamic simulation. Energy, 275, 127477.
Long, J., Shelhamer, E., & Darrell, T. (2015). Fully convolutional networks for semantic
segmentation. Proceedings of the IEEE conference on computer vision and pattern
recognition, Boston, Massachusetts, USA.
Lu, W. (2019). Big data analytics to identify illegal construction waste dumping: A Hong
Kong study. Resources, conservation and recycling, 141, 264-272.
Lu, W., & Chen, J. (2022). Computer vision for solid waste sorting: A critical review of
academic research. Waste Management, 142, 29-43.
https://doi.org/10.1016/j.wasman.2022.02.009
16

645

650

655

660

665

670

675

680

685

690

Lu, W., Chen, J., & Xue, F. (2022). Using computer vision to recognize composition of
construction waste mixtures: A semantic segmentation approach. Resources,
conservation and recycling, 178, 106022.
Lu, W., Yuan, L., & Lee, W. M. (2022). Understanding loading patterns of construction waste
hauling trucks: triangulation between big quantitative and informative qualitative data.
Environmental Science and Pollution Research, 29(33), 50867-50880.
Lu, W., Yuan, L., & Xue, F. (2021). Investigating the bulk density of construction waste: A
big data-driven approach. Resources, conservation and recycling, 169, 105480.
Mishra, P., & Sarawadekar, K. (2019). Polynomial learning rate policy with warm restart for
deep neural network. TENCON 2019-2019 IEEE Region 10 Conference (TENCON),
Kochi, India.
Na, S., Heo, S., Han, S., Shin, Y., & Lee, M. (2022). Development of an artificial intelligence
model to recognise construction waste by applying image data augmentation and
transfer learning. Buildings, 12(2), 175.
Nežerka, V., Zbíral, T., & Trejbal, J. (2024). Machine-learning-assisted classification of
construction and demolition waste fragments using computer vision: Convolution
versus extraction of selected features. Expert Systems with Applications, 238, 121568.
Ni, C., Wu, J., Wang, H., Lu, W., & Zhang, C. (2024). Enhancing Cloud-Based Large
Language Model Processing with Elasticsearch and Transformer Models. arXiv
preprint arXiv:2403.00807.
OpenMMLab. (2020). MMSegmentation: Openmmlab semantic segmentation toolbox and
benchmark. Retrieved 13rd May from https://github.com/openmmlab/mmsegmentation
Ren, T., Liu, S., Zeng, A., Lin, J., Li, K., Cao, H., Chen, J., Huang, X., Chen, Y., & Yan, F.
(2024). Grounded sam: Assembling open-world models for diverse visual tasks. arXiv
preprint arXiv:2401.14159.
Tsipras, D., Santurkar, S., Engstrom, L., Turner, A., & Madry, A. (2018). Robustness may be
at odds with accuracy. arXiv preprint arXiv:1805.12152.
USEPA. (2023). Sustainable Management of Construction and Demolition Materials.
Retrieved 1 December 2023 from https://shorturl.at/hitV9
Wang, X., Girshick, R., Gupta, A., & He, K. (2018). Non-local neural networks. Proceedings
of the IEEE conference on computer vision and pattern recognition, Salt Lake City,
UT, USA.
Wu, H., Zhang, J., Huang, K., Liang, K., & Yu, Y. (2019). Fastfcn: Rethinking dilated
convolution in the backbone for semantic segmentation. arXiv preprint
arXiv:1903.11816.
Xiao, T., Liu, Y., Zhou, B., Jiang, Y., & Sun, J. (2018). Unified perceptual parsing for scene
understanding. Proceedings of the European conference on computer vision (ECCV),
Munich, Germany.
Xu, M., Zhang, Z., Wei, F., Hu, H., & Bai, X. (2023). Side adapter network for openvocabulary semantic segmentation. Proceedings of the IEEE/CVF Conference on
Computer Vision and Pattern Recognition, Vancouver, BC, Canada.
Yong, Q., Wu, H., Wang, J., Chen, R., Yu, B., Zuo, J., & Du, L. (2023). Automatic
identification of illegal construction and demolition waste landfills: A computer vision
approach. Waste Management, 172, 267-277.
Yuan, L., Lu, W., & Xue, F. (2021). Estimation of construction waste composition based on
bulk density: A big data-probability (BD-P) model. Journal of environmental
management, 292, 112822.

17

695

700

705

710

Yuan, L., Yang, B., Lu, W., & Peng, Z. (2024). Carbon footprint accounting across the
construction waste lifecycle: A critical review of research. Environmental Impact
Assessment Review, 107, 107551.
Yuan, Y., Chen, X., & Wang, J. (2020, August 23–28, 2020). Object-contextual
representations for semantic segmentation. Computer Vision–ECCV 2020: 16th
European Conference, Proceedings, Part VI 16, Glasgow, UK.
Zhang, H., Dana, K., Shi, J., Zhang, Z., Wang, X., Tyagi, A., & Agrawal, A. (2018). Context
encoding for semantic segmentation. Proceedings of the IEEE conference on
Computer Vision and Pattern Recognition, Salt Lake City, UT, USA.
Zhao, H., Shi, J., Qi, X., Wang, X., & Jia, J. (2017). Pyramid scene parsing network.
Proceedings of the IEEE conference on computer vision and pattern recognition,
Honolulu, HI, USA.
Zhao, H., Zhang, Y., Liu, S., Shi, J., Loy, C. C., Lin, D., & Jia, J. (2018). Psanet: Point-wise
spatial attention network for scene parsing. Proceedings of the European conference
on computer vision (ECCV), Munich, Germany.
Zhao, Q., Gao, W., Su, Y., Wang, T., & Wang, J. (2023). How can C&D waste recycling do a
carbon emission contribution for construction industry in Japan city? Energy and
Buildings, 298, 113538. https://doi.org/https://doi.org/10.1016/j.enbuild.2023.113538
Zhou, Q., Liu, H., Qiu, Y., & Zheng, W. (2022). Object detection for construction waste based
on an improved YOLOv5 model. Sustainability, 15(1), 681.
https://doi.org/https://doi.org/10.3390/su15010681

18