A Re-Ranking Method Using K-Nearest Weighted Fusion for Person
Re-Identification*
Quang-Huy Che
1,2
, Le-Chuong Nguyen
1,2
, Gia-Nghia Tran
1,2
, Dinh-Duy Phan
1,2
and
Vinh-Tiep Nguyen
1,2
1
University of Information Technology, Ho Chi Minh City, Vietnam
2
Vietnam National University, Ho Chi Minh City, Vietnam
Keywords:
Re-Ranking, Person Re-Identification, Multi-View Fusion, K-Nearest Weighted Fusion.
Abstract:
In person re-identification, re-ranking is a crucial step to enhance the overall accuracy by refining the initial
ranking of retrieved results. Previous studies have mainly focused on features from single-view images, which
can cause view bias and issues like pose variation, viewpoint changes, and occlusions. Using multi-view
features to present a person can help reduce view bias. In this work, we present an efficient re-ranking method
that generates multi-view features by aggregating neighbors’ features using K-nearest Weighted Fusion (KWF)
method. Specifically, we hypothesize that features extracted from re-identification models are highly similar
when representing the same identity. Thus, we select K neighboring features in an unsupervised manner to
generate multi-view features. Additionally, this study explores the weight selection strategies during feature
aggregation, allowing us to identify an effective strategy. Our re-ranking approach does not require model
fine-tuning or extra annotations, making it applicable to large-scale datasets. We evaluate our method on the
person re-identification datasets Market1501, MSMT17, and Occluded-DukeMTMC. The results show that
our method significantly improves Rank@1 and mAP when re-ranking the top M candidates from the initial
ranking results. Specifically, compared to the initial results, our re-ranking method achieves improvements
of 9.8%/22.0% in Rank@1 on the challenging datasets: MSMT17 and Occluded-DukeMTMC, respectively.
Furthermore, our approach demonstrates substantial enhancements in computational efficiency compared to
other re-ranking methods.
1 INTRODUCTION
Person re-identification (ReID) (Luo et al., 2019b;
Wieczorek et al., 2021; Ni et al., 2023; Chen et al.,
2023; Somers et al., 2023; Li et al., 2023) is a com-
puter vision task that involves recognizing and match-
ing a person across multiple images or video frames
from different cameras. The goal is to reidentify a
person despite variations in pose, lighting, camera
views, and occlusion. In a typical ReID system, given
a query image, the system retrieves and ranks candi-
date images from a gallery based on their similarity
to the query. Initial rankings are based on features
from deep learning models and distance metrics. Re-
ranking refines this list to improve re-identification
accuracy, aiming to give higher ranks to relevant im-
ages.
Re-ranking is a powerful technique widely em-
∗
First two authors contribute equally and the fifth is
the corresponding author.
GPU Memory (GB)
4 8 12 16
96.2
94.7
AQE
GNN
20
95.2
Rank@1 (%)
95.7
0
k-reciprocal
KWF (ours)
GCR
-QE
4 8 12 16
70
58
AQE
20
62
Rank@1 (%)
66
0
k-reciprocal
GCR
-QE
GNN
Market1501
Occluded-DukeMTMC
KWF (ours)
Figure 1: Comparing the computational cost and Rank@1
performance of various reranking methods on the Mar-
ket1501 and Occluded-DukeMTMC datasets. The y-axis
shows Rank@1, the x-axis represents GPU memory usage,
and the circle size indicates evaluation time, with larger cir-
cles representing longer evaluation times.
Che, Q.-H., Nguyen, L.-C., Tran, G.-N., Phan, D.-D. and Nguyen, V.-T.
A Re-Ranking Method Using K-Nearest Weighted Fusion for Person Re-Identification.
DOI: 10.5220/0013176100003905
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 14th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2025), pages 79-90
ISBN: 978-989-758-730-6; ISSN: 2184-4313
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.
79

Market1501MSMT17
Occluded-DukeMTMC
Figure 2: Person re-identification datasets often include eight images of the same person taken from different viewpoints
and cameras. The Market1501 and MSMT17 datasets are known for their diverse viewpoints, lighting conditions, and back-
grounds. On the other hand, the Occluded-DukeMTMC dataset poses extra challenges with its complex environments and
frequent partial occlusions, complicating the re-identification process.
ployed in various tasks, including re-identification
(Zhong et al., 2017; Yu et al., 2017; Liu et al., 2018;
Wu et al., 2018; Liu et al., 2020; Zhang et al., 2020).
A notable advantage of many re-ranking techniques
is their ability to be deployed without needing ad-
ditional training samples, allowing them to be ap-
plied to any initial ranking results. Traditionally, com-
mon methods for re-ranking include recalculating im-
age similarities by integrating additional information
and using advanced similarity metrics (Hirzer et al.,
2012; Liao et al., 2015; Yan et al., 2015; Hai Phan,
2022), feature similarity-based methods (Radenovi
´
c
et al., 2018; Chum et al., 2007), neighbor similar-
ity (Bai and Bai, 2016; Zhong et al., 2017; Sarfraz
et al., 2018a; Shen et al., 2012; Qin et al., 2011), and
graph-based approaches (Zhang et al., 2020; Zhang
et al., 2023). However, previous methods still rely on
single-view features directly extracted from the fea-
ture extraction model to represent gallery features.
Relying only on single-view features can result in
a lack of the necessary information to fully and ac-
curately represent pedestrians, especially in complex
situations where images are captured from different
cameras.
The study (Xu et al., 2021; Che et al., 2025) em-
phasized the limitation of view bias in single-view
features when re-identifying pedestrians from differ-
ent camera viewpoints. Since a pedestrian can be cap-
tured by multiple cameras, two images of the same
person may not have the same details, as one image
might lack information present in the other, as illus-
trated in Figure 2. Generating features that capture
information from multiple viewpoints is an effective
solution for mitigating view bias. In this paper, we
approach transforming single-view features to multi-
view features. We draw several conclusions, includ-
ing: (1) single-view features can be effectively trans-
formed into multi-view features, (2) selecting single-
view features can be done effectively in an unsuper-
vised manner if the correct number of K is chosen,
without the need for model fine-tuning, and (3) gen-
erating single-view features by selecting appropriate
weights allows the generation of multi-view features
that better capture the information from single-view
features.
In this paper, we propose generating multi-view
features to represent all images in the gallery set dur-
ing the re-ranking stage, thereby addressing the issue
of view bias. To achieve this, we generate multi-view
features using the K-nearest Weighted Fusion (KWF)
method, which generates multi-view features from
K-nearest neighbor features. Our proposed unsuper-
vised neighbor feature selection method does not re-
quire fine-tuning of pre-trained models. The contri-
butions of this paper are summarized as follows:
• We propose two-stage hierarchical person re-
identification: the first stage involves ranking
based on single-view features, followed by a sec-
ond stage re-ranking using multi-view features.
• We propose the KWF method, a multi-view fea-
ture representation used during the re-ranking
stage. This representation prevents the view bias
problem in person re-identification.
• We study the effectiveness of weight selection
strategies in multi-view feature generation for
KWF, including Uniform, Inverse Distance Power
and Exponential Decay.
• We perform extensive experiments comparing
our proposed methods with other re-ranking ap-
proaches on datasets like Market-1501, MSMT17,
and Occluded-DukeMTMC.
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
80

Gallery images
KWF
Top M candidates
Stage 1 results
Query image
Query image
Top M candidates
Stage 2 results
Multi-view features
Single-view features
KWF KWF KWF KWF
K nearest neighbors
Figure 3: Our person identification pipeline consists of two stages. In the first stage, gallery images are ranked according
to the cosine distance between the query and gallery images. In second stage, the top 5 candidates from the first stage are
re-ranked using multi-view features.
2 RELATED WORK
2.1 Re-Ranking Approach
Image retrieval in general, and person re-
identification in particular, involves searching
for images from large databases based on a query
image. Re-ranking techniques are crucial in refin-
ing the initial search results to enhance retrieval
accuracy. Among these, feature similarity-based
methods (Chum et al., 2007; Radenovi
´
c et al.,
2018; Arandjelovi
´
c and Zisserman, 2012) leverage
the nearest samples with similar features to enrich
the query features by aggregating the features of
neighbors. Additionally, neighbor similarity-based
methods (Bai and Bai, 2016; Zhong et al., 2017; Qin
et al., 2011) rely on the number of shared neighbors
between images, using k-reciprocal nearest neighbors
to efficiently exploit the relationships among im-
ages. Furthermore, graph-based re-ranking methods
(Zhang et al., 2023; Zhang et al., 2020) capture the
topological structure of the data and refine the learned
features, yielding promising results. While some
re-ranking methods do not require extra annotations,
others require human interaction or label supervision
(Bai et al., 2010; Liu et al., 2013; Leng et al., 2015).
Most re-ranking methods focus on directly exploiting
relationships between the initial features. However,
the effectiveness of enhancing multi-view features in
this context remains underexplored. Therefore, in this
study, we propose a method that employs multi-view
features during re-ranking without requiring model
fine-tuning or extra annotations.
Single-view feature K-nearest features Multi-view feature
Figure 4: The process of generating multi-view features
from single-view features: First, the single-view feature cal-
culates the distance to all features in the gallery to find the
K nearest neighbors. Then, the KWF method aggregates the
neighboring features to generate a multi-view feature.
2.2 Multi-View Feature Representation
Previous studies (Wieczorek et al., 2021; Xu et al.,
2021) have explored feature fusion in re-identification
tasks. In (Wieczorek et al., 2021), the authors sug-
gested aggregating features of the same class during
the query process, which proved effective in both ac-
curacy and query time. However, this approach re-
quires additional labels to select representative fea-
tures for images of the same class. Therefore, this
method requires class-level labeling for images when
A Re-Ranking Method Using K-Nearest Weighted Fusion for Person Re-Identification
81

applied in real-world scenarios, highlighting a limi-
tation of this approach. Yinsong et al. (Xu et al.,
2021) proposed feature aggregation using a Graph
Neural Network (GNN). While promising, GNN-
based methods have drawbacks, such as hardware-
dependent computation costs and significant compu-
tational and storage demands for large graphs. This
study introduces a weighted aggregation method for
single-view features based on neighboring features.
Our method selects neighboring features unsupervis-
edly, ensuring efficiency by leveraging the feature
extraction capabilities of pre-trained re-identification
models.
3 METHOD
In this section, we propose a two-stage hierarchical
approach for the person re-identification task. In the
first stage, images in the gallery are ranked based on
their pairwise distance to a given query. To compute
this pairwise distance, we use the cosine distance of
features extracted by a re-identification model, where
the features representing the gallery images are re-
ferred to as single-view features. In the second stage,
the single-view features extracted from the first stage
are transformed using the KWF method. This method
aggregates the neighboring features of the single-view
features to generate multi-view features. To effec-
tively represent multi-view features, we propose effi-
cient weight selection strategies that allow the multi-
view features to capture information from the single-
view features more effectively. Figure 3 illustrates the
two-stage re-identification process.
3.1 Two-Stage Hierarchical Person
Re-Identification
To query an image q within a gallery set of N im-
ages G = {g
i
}
i=1...N
, the distance between q and each
gallery image g
i
∈ G is computed using their fea-
ture vectors. To capture essential visual informa-
tion, we use a pretrained F (·) to extract representa-
tion features. By calculating the distances between
F (q) and each F (g
i
), an initial ranked list R (q, G)
= {g
0
i
}
i=1...N
is generated, where d

F (q),F (g
0
i
)

<
d

F (q),F (g
0
i+1
)

. Our goal is to re-rank R (q,G)
so that more positive samples rank higher in the
list, thereby improving the performance of person re-
identification.
We propose a two-stage hierarchical person re-
identification process as follows:
• Stage 1: Person re-identification ranking involves
ranking gallery images based on the pairwise co-
sine distances between the query and gallery im-
ages in the feature space of a pre-trained feature
extraction model.
• Stage 2: We perform re-ranking of the top M can-
didates from Stage 1 results by computing fea-
ture distances based on multi-view features for the
gallery images. Multi-view features is generated
by K-nearest Weighted Fusion proposed method.
Overall, our approach compares person images
through two hierarchical stages (as depicted in Figure
3): single-view feature ranking stage and multi-view
feature re-ranking stage.
3.2 K-Nearest Weighted Fusion for
Multi-View Features
3.2.1 Generate Multi-View Features from
K-Nearest Neighbors
Instead of representing each image using single-view
features, we propose using multi-view features to rep-
resent the images in the top M candidates from the
initial ranked list. To generate multi-view features
from single-view features, based on the methodology
presented in (Che et al., 2025), we propose the K-
Nearest Weighted Fusion (KWF) method, which ag-
gregates the K-nearest neighboring features for each
single-view feature. Our KWF method effectively
combines neighboring features to mitigate view bias,
providing a more accurate representation of individ-
uals. Specifically, given the results from Stage 1
- R (q,G), we generate a list of the images in the
top M candidates. This list of features is denoted
as N (q, M) = {g
0
j
}
j=1...M
, where each feature of
an image in N (q, M) is single-view feature. With
each single-view feature, K neighboring features are
selected and aggregated into a multi-view feature
through the proposed KWF method. This process is
visualized in Figure 4. In our proposed KWF method,
the nearest neighboring features are selected unsuper-
vised. This selection method may result in neigh-
boring feature lists that include images from differ-
ent classes with the single-view feature, potentially
leading to multi-view features containing inaccurate
information. However, theoretically, the feature ex-
traction model F (·) is trained such that images from
the same class have high similarity, and vice versa.
Thus, we expect that in the nearest neighbor fea-
ture list, the number of images from the same class
as the single-view feature’s image outnumber those
from different classes, allowing positive features to
outweigh and suppress negative ones. By recalculat-
ing the distance between F (q) and each multi-view
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
82

(a) Uniform. (b) Inverse Distance Power (p = 1). (c) Inverse Distance Power (p = 2). (d) Exponential Decay.
Figure 5: The process involves transforming single-view features into multi-view features using the KWF method. In each
figure, the red star represents the single-view feature, gray circles represent the nearest neighboring features (with K = 6), and
the green quadrilateral represents the generated multi-view feature.
Table 1: Comparison of our proposed method with various approaches on the Market-1501, MSMT17, and Occluded-
DukeMTMC datasets. The results are reported in terms of Rank@1 and mAP. The best results are marked in bold, and
the second-best results are indicated with underlining.
Market1501 MSMT17 Occluded-DukeMTMC
Rank@1 mAP Rank@1 mAP Rank@1 mAP
BoT Baseline 94.5 85.9 74.1 50.2 48.7 42.6
AQE (Chum et al., 2007) 94.7 91.2 69.8 55.7 58.3 55.7
α-QE (Radenovi
´
c et al., 2018) 95.4 92.0 70.2 55.9 57.9 55.6
k-reciprocal (Zhong et al., 2017) 95.4 94.2 79.8 66.8 58.5 60.3
ECN (Sarfraz et al., 2018b) 95.8 93.2 - - - -
LBR (Luo et al., 2019a) 95.7 91.3 - - - -
GNN (Zhang et al., 2020) 95.8 94.6 - - 59.3 61.5
GCR (Zhang et al., 2023) 96.1 94.7 - - 63.6 64.3
Ours
Uniform 95.4 90.6 82.4 60.0 65.2 54.8
Inverse Distance Power (p = 2) 96.2 87.3 83.9 56.1 70.7 52.2
Exponential Decay 96.1 90.8 83.2 60.3 66.9 55.3
feature KW F

F (g
j
)

with g
j
∈ N (q,M), we ob-
tain N
∗
(q,M) = {g
∗
j
}
j=1...M
. By re-ranking based
on multi-view features, N
∗
(q,M) have more positive
samples ranked higher compared to N (q, M). To en-
sure cross-view matching settings, which is crucial in
the re-identification task, the KWF feature aggrega-
tion method excludes images in the K-nearest neigh-
bors list with the same camera ID as the query image.
It is important to note that all features used in Stage
2 are derived from Stage 1 without the need for re-
extraction.
3.2.2 Weight Selection Strategy
When generating multi-view features, we aggregate
single-view features using the KWF method. In the
study by (Wieczorek et al., 2021), the aggregated fea-
ture of an image is computed by averaging the single-
view features. This approach works well when the
images used for aggregation belong to the same class
as the single-view feature. However, in our unsu-
pervised selection method, treating the contributions
of neighboring features equally may lead to subopti-
mal results. Specifically, features with high similar-
ity (mainly from the same class) and those with low
similarity (possibly from different classes) contribute
equally to the multi-view feature. Therefore, apply-
ing weights during the aggregation process becomes
a crucial step. In this study, we thoroughly explore
different weight selection strategies for aggregation,
as illustrated in Figure 5. Each weight selection strat-
egy generates distinct multi-view features, resulting
in unique aggregated representations. The visualized
results demonstrate the impact of these weight selec-
tion strategies. With f = F (I) as the feature extracted
by the model for any given image I, the multi-view
feature f
(mv)
can be generated following:
f
(mv)
=
K
∑
k=1
w
k
·
·
· f
(nn)
k
(1)
• f
(mv)
: Multi-view feature representing the image
I
• f
(nn)
k
: The k-th neighboring feature in the list of
A Re-Ranking Method Using K-Nearest Weighted Fusion for Person Re-Identification
83

the K nearest neighbors of f
• w
k
: The aggregation weight corresponding to
f
(nn)
k
We are suggesting various methods for weight se-
lection strategies to enhance effectiveness:
Uniform Weighting: This method assigns equal
weights to all K nearest neighbor features, regardless
of their distance from the single-view feature. Each
neighbor has the same influence on generating the
multi-view feature:
w
k
=
1
K
(2)
Inverse Distance Power Weighting: In this
method, weights are assigned inversely proportional
to the p-th power of the distance between each
neighboring feature f
(nn)
k
and the single-view feature
f . The parameter p controls the rate at which
weights decrease as the distance increases. A larger
p emphasizes closer features more strongly, while a
smaller p allows for a more balanced contribution
from distant features. The weights are calculated as:
w
k
=
1/d
p

f , f
(nn)
k

∑
K
j=1
1/d
p

f , f
(nn)
j

(3)
After extensive experimentation, we selected p =
2 as our default value.
Exponential Decay Weighting: This method em-
ploys an exponential function to assign weights based
on distance. The weights decrease exponentially as
the distance increases, ensuring that only the closest
neighbors significantly influence the multi-view fea-
ture:
w
k
=
e
−d

f , f
(nn)
k

∑
K
j=1
e
−d

f , f
(nn)
j

(4)
4 RESULTS
4.1 Datasets and Settings
Our main results are trained and evaluated on three-
person re-identification datasets:
• Market-1501 (Zheng et al., 2015): The Market-
1501 dataset, gathered at Tsinghua University,
features 32,668 images of 1,501 individuals cap-
tured by six cameras positioned outside a super-
market. The dataset is split into 12,936 training
images from 751 individuals and 19,732 testing
images from 750 individuals, with 3,368 of the
test images designated as the query set.
• MSMT17 (Wei et al., 2018): MSMT17 is an
extensive dataset for multi-scene and multi-time
person re-identification. It comprises 180 hours
of video from 12 outdoor and 3 indoor cam-
eras across 12 different time slots. The dataset
includes 4,101 annotated identities and 126,441
bounding boxes, providing a robust foundation for
re-identification studies.
• Occluded-DukeMTMC (Miao et al., 2019): This
dataset contains 15,618 training images, 17,661
gallery images, and 2,210 occluded query images.
Our experiments on the Occluded-DukeMTMC
dataset demonstrate the efficacy of our method in
handling occluded person re-identification. Im-
portantly, our approach does not require manual
cropping during preprocessing.
We utilize the BoT (Luo et al., 2019b) with a
ResNet-50 backbone to extract features while evalu-
ating our proposed. We use two evaluation metrics
to assess the performance of person Re-ID methods
across all datasets. The first metric is the Cumulative
Matching Characteristic (CMC). Viewing re-ID as a
ranking problem, we report the cumulative matching
accuracy at Rank@1. The second metric is the mean
Average Precision (mAP). We evaluate the proposed
re-ranking method on the top 100 candidates in the
initial result list (M = 100). For K, which is the
number of nearest neighbors, we select K = 6 for the
MSMT17 and Occluded-DukeMTMC datasets, while
for Market1501, we choose K = 4. The selection of
these hyperparameters M and K is further analyzed in
Section 4.3. All experiments used the PyTorch frame-
work on an GeForce RTX 4090 GPU with 32GB of
RAM and an Intel(R) Core(TM) i9-10900X proces-
sor.
4.2 Main Results
4.2.1 Quantitative Results
As shown in Table 1, our method consistently out-
performs the traditional Stage 1 alone in both
Rank@1 and mAP metrics, regardless of the weight
selection strategies used. Among the three strate-
gies, Inverse Distance Power Weighting (p = 2)
demonstrates the best Rank@1 performance, while
Exponential Decay Weighting shows uniform im-
provement in both Rank@1 and mAP. Specifically,
with Inverse Distance Power Weighting, our method
achieves improvements of 1.7%, 9.8%, and 22.0%
in Rank@1 on the Market-1501, MSMT17, and
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
84

Table 2: Comparison of the computation time and memory usage in re-ranking methods.
Method
Market1501 Occluded-DukeMTMC
Evaluate time (↓) GPU Memory (↓) Evaluate time (↓) GPU Memory (↓)
AQE (Chum et al., 2007) 007.8s 10.55GB 008.8s 12.53GB
α-QE (Radenovi
´
c et al., 2018) 007.9s 10.55GB 008.8s 12.53GB
k-reciprocal (Zhong et al., 2017) 146.0s 05.63GB 150.8s 05.62GB
GNN (Zhang et al., 2020) 008.6s 04.75GB 10.97s 04.96GB
GCR (Zhang et al., 2023) 034.5s 20.97GB 035.7s 23.24GB
Our 008.5s 01.16GB 006.1s 01.11GB
Table 3: Performance of our method on Market-1501 and Occluded-DukeMTMC across different top M candidates.
Top-M
Market-1501 Occluded-DukeMTMC
Rank-1 mAP Time(ms) Rank-1 mAP Time(ms)
20 96.0 88.3 1.41 58.9 46.7 1.58
40 96.1 89.8 1.47 62.6 50.6 1.70
60 96.1 90.3 1.59 64.7 52.8 1.73
80 96.1 90.6 1.70 66.1 54.2 1.85
100 96.1 90.8 1.76 66.9 55.3 1.93
120 96.1 90.9 1.88 67.6 56.1 2.03
140 96.1 90.9 1.94 68.1 56.8 2.17
160 96.1 91.0 2.03 68.6 57.4 2.25
Table 4: Performance of our proposed on Market-1501 and
Occluded-DukeMTMC across different K values.
K
Market-1501 Occluded-DukeMTMC
Rank-1 mAP Rank-1 mAP
2 95.9 89.3 69.3 53.2
3 95.9 90.2 68.1 54.3
4 96.1 90.8 67.8 54.9
5 95.7 91.0 66.9 55.2
6 95.8 91.1 66.9 55.3
7 95.7 91.0 65.8 55.0
8 95.4 90.9 65.5 54.7
9 95.2 90.7 64.3 54.3
10 95.2 90.6 63.8 53.9
Occluded-DukeMTMC datasets, respectively. More-
over, when comparing our method to other post-
processing methods, it delivers higher Rank@1 ac-
curacy, particularly on the MSMT17 and Occluded-
DukeMTMC datasets. Our approach focuses on re-
ranking the top M candidates based on multi-view
features, making it particularly effective when eval-
uated on the Rank@1 metric. Our method improves
person re-identification, especially in datasets with
occlusions and complex variations, such as MSMT17
and Occluded-DukeMTMC. The proposed weight
selection strategies enhance feature representation,
leading to better identity discrimination.
To compare the computational cost of our method
with other approaches, we evaluate memory usage
and evaluation time (measuring the time five times
and taking the average result). All methods are imple-
mented on a GPU (except for the k-reciprocal (Zhong
et al., 2017), executed on the GPU in Stage 1 and
the CPU in Stage 2). We do not account for the im-
age feature extraction time; instead, we only measure
each method’s time to process queries and perform re-
ranking. The results in Table 2 demonstrate that our
method requires significantly less memory than other
methods, utilizing only about 1GB of GPU mem-
ory. Our approach involves only the time needed to
generate multi-view features without requiring addi-
tional memory to store new features or auxiliary com-
ponents. Additionally, with competitive processing
time, our method shows strong applicability for large-
scale retrieval systems.
4.2.2 Qualitative Results
Figure 6 visualizes some results after using our re-
ranking method. Stage 1 shows the initial results
based on direct feature distance computation, while
Stage 2 presents the re-ranking results using multi-
view features. The outcomes demonstrate the ro-
bustness of our method in re-identifying images with
changes in viewpoint and background across the Mar-
ket1501 and MSMT17 datasets. Even more im-
pressively, on the Occluded-DukeMTMC dataset, our
re-identification approach accurately handles cases
A Re-Ranking Method Using K-Nearest Weighted Fusion for Person Re-Identification
85

d
Query
Rank@10
Stage 1
Stage 2
Query
Rank@10
(a) Market1501.
Query
Rank@10
Query
Rank@10
Stage 1
Stage 2
(b) MSMT17.
Query
Rank@10
Query
Rank@10
Stage 1
Stage 2
(c) Occluded-DukeMTMC.
Figure 6: The results from the experiments demonstrate the effectiveness of the proposed reranking method. In each sub-
figure (Market1501, MSMT17, and Occluded-DukeMTMC), the query images are on the left, followed by columns showing
the top 10 most retrieved results. The results from Stage 1 are on the top row of each dataset, while the reranking results from
Stage 2 are on the bottom row. Blue and red boxes indicate true positives and false positives, respectively.
where the persons are occluded. These results show
that our method effectively aggregates information,
significantly enhancing retrieval performance during
re-ranking.
4.3 Ablation Studies and Analysis
In this section, we perform extensive ablation stud-
ies on the Market1501 and Occluded-DukeMTMC
datasets to examine the impact and sensitivity of var-
ious hyperparameters.
4.3.1 Effect of Different K
To analyze the impact of the number of nearest neigh-
bors K we conducted experiments on the Market-
1501 and Occluded-DukeMTMC datasets. Table
4 and Figure 8 demonstrate the trade-off between
Rank@1 and mAP as K increases from 2 to 10.
Specifically, Rank@1 tends to decrease as K in-
creases, while mAP rises to K = 6 , after which
it starts to decline. For the Occluded-DukeMTMC
dataset, we chose K = 6 as it provides the best bal-
Table 5: Performance of our proposed on Market-1501 and
Occluded DukeMTMC across different α values.
α
α
α
Market-1501 Occluded-DukeMTMC
Rank-1 mAP Rank-1 mAP
0.0 94.5 85.9 48.7 42.6
0.2 95.0 87.7 53.7 47.0
0.4 95.7 89.1 58.1 50.9
0.6 95.9 90.0 63.0 53.6
0.8 96.0 90.5 66.0 54.9
1.0 96.1 90.8 66.9 55.3
ance between Rank-1 and mAP. For the Market-1501
dataset, we selected K = 4 to maintain Rank-1 perfor-
mance, as this metric decreases rapidly with increas-
ing K.
4.3.2 Effect of Different Top M Candidates
The selection of the top M candidates, as observed in
Table 3 and Figure 7, illustrates the trade-off between
accuracy and query time. The query time refers to the
duration required to perform a single query, measured
five times, with the average value reported. Choosing
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
86

(a) Market-1501. (b) Occluded-DukeMTMC.
Figure 7: Results when select different top M candidates.
(a) Market-1501. (b) Occluded-DukeMTMC.
Figure 8: Results when changing K for finding nearest neighbors.
(a) Market-1501. (b) Occluded-DukeMTMC.
Figure 9: Results when sweeping across α for linearly combining multi-view and single-view features.
a higher M value means re-ranking more candidates
using multi-view features, which improves accuracy
but also requires more time for feature aggregation.
Therefore, we choose M = 100 in this study to balance
accuracy and query time.
4.3.3 Combining Multi-View and Single-View
Features
In this section, we explore combining the generated
multi-view feature and the original single-view fea-
ture by linearly combining f and f
(mv)
using α ∈ {0.0,
A Re-Ranking Method Using K-Nearest Weighted Fusion for Person Re-Identification
87

Table 6: The results of our method applied to different baselines.
Method
Market1501 Occluded-DukeMTMC MSMT17
Rank@1 mAP Rank@1 mAP Rank@1 mAP
CLIP-ReID (CNN) 94.7 88.1 54.2 47.4 - -
+ Ours 95.7 (↑ 1.0) 91.8 (↑ 3.7) 68.1 (↑ 13.9) 59.5 (↑ 12.1) - -
CLIP-ReID (ViT) 93.3 86.4 60.8 53.5 - -
+ Ours 93.7 (↑ 0.4) 87.8 (↑ 1.4) 64.8 (↑ 4.0) 60.1 (↑ 6.6) - -
BoT (R101-ibn) 95.4 88.9 - - 81.0 59.4
+ Ours 96.1 (↑ 1.7) 92.2 (↑ 3.3) - - 86.3 (↑ 5.3) 67.5 (↑ 8.1)
Table 7: Experimental results evaluating our method with different indexing techniques, highlighting the trade-offs between
query time and re-identification accuracy.
Market-1501 Occluded-DukeMTMC
Time (s) Time (s)
Type
CPU GPU
Rank@1 mAP
CPU GPU
Rank@1 mAP
Normal 597.21 8.46 96.1 90.8 423.26 6.02 69.1 55.3
IndexIVFPQ
357.84
↓ 40.08%
7.76
↓ 8.27%
95.1
↓ 1.0
88.4
↓ 2.4
237.31
↓ 43.93%
5.11
↓ 15.12%
62.3
↓ 6.8
51.7
↓ 3.6
IndexIVFFlat
354.63
↓ 40.62%
7.19
↓ 15.01%
95.5
↓ 0.6
89.0
↓ 1.8
232.42
↓ 45.09%
4.74
↓ 21.26%
61.4
↓ 7.7
51.4
↓ 3.9
IndexLSH
515.09
↓ 13.75%
-
96.2
↑ 0.1
88.0
↓ 2.8
358.89
↓ 15.21%
-
70.7
↑ 1.6
53.2
↓ 2.1
0.2, 0.4,..., 1.0}:
f
∗
= (1 − α)× f + α × f
(mv)
(5)
We found that varying α impacts both Rank@1
and mAP results across the Market-1501 and
Occluded-DukeMTMC datasets. The results in Fig-
ure 9 and Table 5 show that the larger the value of
α, the higher the performance. This indicates that the
more significant the contribution of multi-view fea-
tures, the better the re-ranking performance. There-
fore in all experiments we choose α = 1 and ignore
the contribution of f .
4.3.4 Effectiveness in Different Baselines
In addition to the results on the baseline BoT Resnet-
50, we also experimented with our method on dif-
ferent backbones. Table 6 shows the results on var-
ious baselines, including BoT Resnet101-ibn (Luo
et al., 2019b), CLIP-ReID Resnet-50, and ViT-B/16
(Li et al., 2023). The results demonstrate that our
method can be easily applied to other pre-trained
models, leading to significant improvements in both
Rank@1 and mAP.
4.3.5 Time Cost Comparison with Different
Feature Index Methods
In real-world applications such as surveillance and
security, indexing is crucial in managing large-scale
galleries. Most available indexing structures involve
a trade-off between query time and accuracy. There-
fore, in this study, we evaluate our method using vari-
ous indexing types (Johnson et al., 2019; Douze et al.,
2024). Table 7 presents our experimental results.
Across both datasets, using IndexIVFPQ and Index-
IVFFlat demonstrates a similar trade-off: a slight de-
crease in re-identification accuracy but a significant
reduction in evaluation time on the CPU. On the other
hand, when using IndexLSH, the time improvement
is not substantial; however, there is a minor increase
in Rank@1 accuracy compared to the standard ap-
proach. These experiments provide deeper insights
into the practical application of our method in real-
world query systems.
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
88

5 DISCUSSION AND
CONCLUSION
5.1 Limitations
In this study, we primarily focus on exploring the po-
tential of a multi-view feature-based approach for the
re-ranking stage without optimizing the re-ranking
method, resulting in less competitive mAP scores.
Additionally, since our methods select features in an
unsupervised manner, the performance depends on
the pre-trained models. Therefore, the performance
of our method relies on the performance of the pre-
trained person re-identification models. Finally, we
have yet to investigate the effectiveness of multi-view
features in other retrieval tasks, which could be an ex-
citing direction for future research.
5.2 Conclusion
In this study, we proposed a two-stage hierarchical
person re-identification approach combining single-
view and multi-view features. Introducing the K-
nearest Weighted Fusion (KWF) method addressed
the challenges posed by view bias and significantly
improved re-ranking performance without requiring
additional fine-tuning or annotations. Experimental
results on Market-1501, MSMT17, and Occluded-
DukeMTMC datasets demonstrate that our method
outperforms existing re-ranking techniques in Rank-1
accuracy while maintaining computational efficiency.
Our approach improved substantially on challenging
datasets with occlusions, highlighting its robustness
and practical applicability. This work advances the re-
ranking in person re-identification and opens avenues
for future research in adaptive feature aggregation and
the application of multi-view representations to other
domains. By optimizing feature representation, we
aim to contribute to developing more accurate and ef-
ficient retrieval systems for real-world applications.
ACKNOWLEDGMENTS
This research is funded by University of Information
Technology-Vietnam National University of Ho Chi
Minh city under grant number D1-2024-70.
REFERENCES
Arandjelovi
´
c, R. and Zisserman, A. (2012). Three things
everyone should know to improve object retrieval. In
2012 IEEE conference on computer vision and pattern
recognition, pages 2911–2918. IEEE.
Bai, S. and Bai, X. (2016). Sparse contextual activation
for efficient visual re-ranking. IEEE Transactions on
Image Processing, 25(3):1056–1069.
Bai, X., Yang, X., Latecki, L. J., Liu, W., and Tu, Z. (2010).
Learning context-sensitive shape similarity by graph
transduction. IEEE Transactions on Pattern Analysis
and Machine Intelligence, 32(5):861–874.
Che, Q.-H., Nguyen, L.-C., Luu, D.-T., and Nguyen, V.-
T. (2025). Enhancing person re-identification via
uncertainty feature fusion method and auto-weighted
measure combination. Knowledge-Based Systems,
307:112737.
Chen, W., Xu, X., Jia, J., Luo, H., Wang, Y., Wang, F.,
Jin, R., and Sun, X. (2023). Beyond appearance: a
semantic controllable self-supervised learning frame-
work for human-centric visual tasks. In Proceedings
of the IEEE/CVF conference on computer vision and
pattern recognition, pages 15050–15061.
Chum, O., Philbin, J., Sivic, J., Isard, M., and Zisserman, A.
(2007). Total recall: Automatic query expansion with
a generative feature model for object retrieval. In 2007
IEEE 11th International Conference on Computer Vi-
sion, pages 1–8. IEEE.
Douze, M., Guzhva, A., Deng, C., Johnson, J., Szilvasy, G.,
Mazar
´
e, P.-E., Lomeli, M., Hosseini, L., and J
´
egou, H.
(2024). The faiss library.
Hai Phan, A. N. (2022). Deepface-emd: Re-ranking us-
ing patch-wise earth mover’s distance improves out-
of-distribution face identification. In Proceedings of
the IEEE/CVF Conference on Computer Vision and
Pattern Recognition.
Hirzer, M., Roth, P. M., K
¨
ostinger, M., and Bischof, H.
(2012). Relaxed pairwise learned metric for person re-
identification. In Computer Vision–ECCV 2012: 12th
European Conference on Computer Vision, Florence,
Italy, October 7-13, 2012, Proceedings, Part VI 12,
pages 780–793. Springer.
Johnson, J., Douze, M., and J
´
egou, H. (2019). Billion-scale
similarity search with GPUs. IEEE Transactions on
Big Data, 7(3):535–547.
Leng, Q., Hu, R., Liang, C., Wang, Y., and Chen, J. (2015).
Person re-identification with content and context re-
ranking. Multimedia Tools Appl., 74(17):6989–7014.
Li, S., Sun, L., and Li, Q. (2023). Clip-reid: Exploit-
ing vision-language model for image re-identification
without concrete text labels. Proceedings of the AAAI
Conference on Artificial Intelligence, 37:1405–1413.
Liao, S., Hu, Y., Zhu, X., and Li, S. Z. (2015). Person re-
identification by local maximal occurrence represen-
tation and metric learning. In Proceedings of the IEEE
conference on computer vision and pattern recogni-
tion, pages 2197–2206.
Liu, C., Loy, C. C., Gong, S., and Wang, G. (2013). Pop:
Person re-identification post-rank optimisation. In
2013 IEEE International Conference on Computer Vi-
sion, pages 441–448.
Liu, Y., Shang, L., and Song, A. (2018). Adaptive re-
A Re-Ranking Method Using K-Nearest Weighted Fusion for Person Re-Identification
89

ranking of deep feature for person re-identification.
arXiv preprint arXiv:1811.08561.
Liu, Y., Shen, J., and He, H. (2020). Multi-attention deep
reinforcement learning and re-ranking for vehicle re-
identification. Neurocomputing, 414:27–35.
Luo, C., Chen, Y., Wang, N., and Zhang, Z.-X.
(2019a). Spectral feature transformation for person
re-identification. In 2019 IEEE/CVF International
Conference on Computer Vision (ICCV), pages 4975–
4984.
Luo, H., Gu, Y., Liao, X., Lai, S., and Jiang, W. (2019b).
Bag of tricks and a strong baseline for deep person re-
identification. In The IEEE Conference on Computer
Vision and Pattern Recognition (CVPR) Workshops.
Miao, J., Wu, Y., Liu, P., Ding, Y., and Yang, Y. (2019).
Pose-guided feature alignment for occluded person re-
identification. In Proceedings of the IEEE/CVF Inter-
national Conference on Computer Vision (ICCV).
Ni, H., Li, Y., Gao, L., Shen, H. T., and Song, J. (2023).
Part-aware transformer for generalizable person re-
identification. In Proceedings of the IEEE/CVF in-
ternational conference on computer vision, pages
11280–11289.
Qin, D., Gammeter, S., Bossard, L., Quack, T., and
Van Gool, L. (2011). Hello neighbor: Accurate ob-
ject retrieval with k-reciprocal nearest neighbors. In
CVPR 2011, pages 777–784. IEEE.
Radenovi
´
c, F., Tolias, G., and Chum, O. (2018). Fine-tuning
cnn image retrieval with no human annotation. IEEE
transactions on pattern analysis and machine intelli-
gence, 41(7):1655–1668.
Sarfraz, M. S., Schumann, A., Eberle, A., and Stiefelhagen,
R. (2018a). A pose-sensitive embedding for person re-
identification with expanded cross neighborhood re-
ranking. In 2018 IEEE/CVF Conference on Computer
Vision and Pattern Recognition, pages 420–429.
Sarfraz, M. S., Schumann, A., Eberle, A., and Stiefelhagen,
R. (2018b). A pose-sensitive embedding for person re-
identification with expanded cross neighborhood re-
ranking. In 2018 IEEE/CVF Conference on Computer
Vision and Pattern Recognition, pages 420–429.
Shen, X., Lin, Z., Brandt, J., Avidan, S., and Wu,
Y. (2012). Object retrieval and localization with
spatially-constrained similarity measure and k-nn re-
ranking. In 2012 IEEE Conference on Computer
Vision and Pattern Recognition, pages 3013–3020.
IEEE.
Somers, V., De Vleeschouwer, C., and Alahi, A. (2023).
Body part-based representation learning for occluded
person re-identification. In Proceedings of the
IEEE/CVF winter conference on applications of com-
puter vision, pages 1613–1623.
Wei, L., Zhang, S., Gao, W., and Tian, Q. (2018). Per-
son transfer gan to bridge domain gap for person
re-identification. In 2018 IEEE/CVF Conference on
Computer Vision and Pattern Recognition, pages 79–
88.
Wieczorek, M., Rychalska, B., and Dabrowski, J. (2021).
On the unreasonable effectiveness of centroids in im-
age retrieval. In Mantoro, T., Lee, M., Ayu, M. A.,
Wong, K. W., and Hidayanto, A. N., editors, Neu-
ral Information Processing, pages 212–223, Cham.
Springer International Publishing.
Wu, F., Yan, S., Smith, J. S., and Zhang, B. (2018). Joint
semi-supervised learning and re-ranking for vehicle
re-identification. In 2018 24th International Confer-
ence on Pattern Recognition (ICPR), pages 278–283.
IEEE.
Xu, Y., Jiang, Z., Men, A., Wang, H., and Luo, H. (2021).
Multi-view feature fusion for person re-identification.
Knowledge-Based Systems, 229:107344.
Yan, C., Shan, S., Wang, D., Li, H., and Chen, X. (2015).
View-adaptive metric learning for multi-view person
re-identification. In Computer Vision–ACCV 2014:
12th Asian Conference on Computer Vision, Singa-
pore, Singapore, November 1-5, 2014, Revised Se-
lected Papers, Part II 12, pages 688–702. Springer.
Yu, R., Zhou, Z., Bai, S., and Bai, X. (2017). Di-
vide and fuse: A re-ranking approach for person re-
identification. arXiv preprint arXiv:1708.04169.
Zhang, X., Jiang, M., Zheng, Z., Tan, X., Ding, E., and
Yang, Y. (2020). Understanding image retrieval re-
ranking: A graph neural network perspective. arXiv
preprint arXiv:2012.07620.
Zhang, Y., Qian, Q., Wang, H., Liu, C., Chen, W., and
Wang, F. (2023). Graph convolution based efficient
re-ranking for visual retrieval. IEEE Transactions on
Multimedia, 26:1089–1101.
Zheng, L., Shen, L., Tian, L., Wang, S., Wang, J., and
Tian, Q. (2015). Scalable person re-identification: A
benchmark. In 2015 IEEE International Conference
on Computer Vision (ICCV), pages 1116–1124.
Zhong, Z., Zheng, L., Cao, D., and Li, S. (2017). Re-
ranking person re-identification with k-reciprocal en-
coding. In 2017 IEEE Conference on Computer Vision
and Pattern Recognition (CVPR), pages 3652–3661.
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
90