https://khawar-islam.github.io/fpvt/

1<!DOCTYPE html
2  PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN" "http://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd">
3<html xmlns="http://www.w3.org/1999/xhtml">
4
5<head>
6  <meta http-equiv="Content-Type" content="text/html; charset=UTF-8">
7  <title>Face Pyramid Vision Transformer</title>
8  <!--=================Meta tags==========================-->
9  <meta name="robots" content="index,follow">
10  <meta name="description"
11    content="Face Pyramid Vision Transformer">
12  <meta name="keywords" content="face recognition, age-invariant face recongition, vision transformer, pyramid vision transformerFPVT">
13  <link rel="author" href="https://khawar-islam.github.io/">
14  <!--=================js==========================-->
15  <link href="./css.css" rel="stylesheet" type="text/css">
16  <link rel="stylesheet" type="text/css" href="main.css" media="screen">
17  

17<script src="./effect.js "></script>

17
18  <!-- Latex -->
19  

19<script type="text/x-mathjax-config">
20    MathJax.Hub.Config({
21        tex2jax: {inlineMath: [['$','$'], ['\\(','\\)']]},
22        TeX: { equationNumbers: { autoNumber: "AMS" } },
23      });
24      </script>

24
25  

25<script type="text/javascript" async
26    src="https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.5/latest.js?config=TeX-MML-AM_CHTML">
27    </script>

27
28  <!--=================Google Analytics==========================-->

28
29  <script async src="https://www.googletagmanager.com/gtag/js?id=

29UA-129775907-1

29"></script>
30  

30<script>
31    

31window.dataLayer = window.dataLayer || [];
32
33    function gtag() {
34      dataLayer.push(arguments);
35    }
36    gtag('js', new Date());
37
38    gtag('config', '

38UA-129775907-1

38');
39  

39</script>

39
40</head>
41
42<body>
43  <div id="content">
44    <div id="content-inner">
45      <div class="section head">
46        <h1>
47          <font color="Tomato">FPVT</font>--Face Pyramid Vision Transformer
48        </h1>
49
50        <div class="authors">
51          <a href="https://khawar-islam.github.io/" target="_blank">Khawar Islam</a> <sup>1</sup>
52          &nbsp;&nbsp;&nbsp;&nbsp;
53          <a href="https://zaighamz.com/" target="_blank">Muhammad Zaigham Zaheer</a> <sup>2,4</sup> &nbsp;&nbsp;&nbsp;&nbsp;
54          <a href="https://itu.edu.pk/faculty-itu/dr-arif-mahmood/" target="_blank">Arif Mahmood</a> <sup>3</sup>
55        </div>
56
57        <div class="affiliations ">
58          <sup>1</sup> FloppyDisk.AI, Pakistan &nbsp;
59          <sup>2</sup> Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE <br>
60          <sup>3</sup> Information Technology University, Pakistan &nbsp;
61          <sup>4</sup> Electronics and Telecommunications Research Institute, South Korea <br>
62        </div>
63
64        <!--=================Conference Name==========================-->
65        <div class="authors">
66          <a href="https://bmvc2022.org/programme/papers/" target="_blank">British Machine Vision Conference (BMVC) 2022</a>
67        </div>
68        <!--=================Conference Name==========================-->
69
70
71        <div class="link">
72          <a href="https://openaccess.thecvf.com/content/CVPR2021W/CLIC/papers/Islam_Image_Compression_With_Recurrent_Neural_Network_and_Generalized_Divisive_Normalization_CVPRW_2021_paper.pdf" target="_blank">[Paper]</a>&nbsp;
73          <a href="https://github.com/khawar-islam/FPVT_BMVC22" target="_blank">[Code]</a>
74          <a href="https://youtu.be/B02dOB0PXxk" target="_blank">[Video]</a>
75          <a href="./bibtex.txt" target="_blank">[Bibtex]</a>
76      
77        </div>
78       
79<!--=================Tabs==========================-->
80        <ul id="tabs">
81          <li><a href="#Highlights" name="#tab1">Highlights</a></li>
82          <li><a href="#Evaluation & Results" name="#tab2">Evaluation & Results</a></li>
83          <li><a href="#Citation" name="#tab3">Citation</a></li>
84          <li><a href="#Contact" name="#tab4">Contact</a></li>
85      </div>
86      <br>
87      <!--=================Teasers==========================-->
88      <div class="section abstract" class="img_container">
89        <center><img src="./fpvt.png" border="0" width="95%"></center>
90        <br>
91        <p style="text-align:justify">
92          A simplified view of Face Pyramid Vision Transformer (FPVT) with limited computational resources. 
93          The whole network represents a pyramid structure, every stage comprises of an improved patch embedding layer 
94          and encoder layer.Following progressive shrinking strategy, the output resolution is diverse at every stage from high to 
95          low resolution.
96        </p>
97      </div>
98      
99       <!--=================News==========================-->
100      <div class="section Highlights", id="Highlights">
101        <h2><font color="red">News</font></h2>
102        <ol>
103          <li>6/10/2022: We will release the extended version of our FPVT paper soon.</li>
104        </ol>
105      </div>
106        
107      <!--=================Highlights==========================-->
108      <div class="section Highlights", id="Highlights">
109        <h2>Highlights</h2>
110        <ol>
111          <li>FPVT introduces a CFFN that extracts locality information</li>
112          <li>We present Face Pyramid Vision Transformer (FPVT) to learn multi-scale discriminative</li>
113          <li>We introduce IPE which utilizes all benefits of CNNs to model low-level edges to higher-level semantic primitives</li>
114        </ol>
115      </div>
116      
117      <!--=================Abstract==========================-->
118      <div class="section abstract">
119        <h2>Abstract</h2>
120        <br>
121        <p style="text-align:justify">
122          A novel Face Pyramid Vision Transformer (FPVT) is proposed to learn a discriminative multi-scale facial representations for face recognition and verification. In FPVT,
123          Face Spatial Reduction Attention (FSRA) and Dimensionality Reduction (F

123DR) layers
124          are employed to reduce the computations by compacting the feature maps. An Improved Patch Embedding (IPE) algorithm is proposed to exploit the benefits of CNNs in
125          ViTs (e.g., shared weights, local context, and receptive fields) to model low-level edges to higher-level semantic primitives. Within FPVT framework, a Convolutional Feed-
126          Forward Network (CFFN) is proposed that extracts locality information to learn low
127          level facial information. The proposed FPVT is evaluated on seven benchmark datasets
128          and compared with ten existing state-of-the-art methods including CNNs, pure ViTs and
129          Convolutional ViTs. Despite fewer parameters, FPVT has demonstrated excellent performance over the compared methods.
130        </p>
131      </div>
132        
133     <!--=================Experiments==========================-->
134      <div class="section Evaluation & Results" , id="Evaluation & Results">
135        <h2>Evaluation & Results</h2>
136        <p style="text-align:justify">
137          We conduct multiple ablation studies to validate the impact of our proposed work in our FPVT modules. 
138          Table. 1 and Table 2 present the results of CNNs, pure ViTs, and Convolutional ViTs on seven datasets. 
139          PVT (referred to as baseline) is a standard pure pyramid transformer without convolution. 
140          We choose PVT as a baseline due to two main reasons i) it generates multi-scale features 
141          ii) the number of parameters is larger than ResNet18. 
142          The introduction of the IPE block leads to a gain of 1% in terms of performance, highlighting the 
143          impact of convolutional tokens. IPE block improves the performance on LFW from 78.8% to 82.9%, 
144          CFP_FF from 75.2% to 85.5%, CFP_FP from 52.9% to 65.6%, AgeDB from 59.9% to 65.6%, CALFW from 
145          66.8% to 70.1%, and CPLFW from 55.1% to 59%.
146        </p>
147        <div id="img_intro_examples" class="img_container">
148          <center><img src="./figs/table2.png" border="0" width="70%"></center>
149          <br>
150        <p style="text-align:justify">
151          Table 1: Evaluation results on CFP_FF, CFP_FP and VGG2_FP. The top first three models
152          lists the results of the CNN models, while the mid-block presents the performance of pure
153          ViTs architectures. The third block shows the results of Conv. ViTs models.
154        </p>
155        </div>
156        <br />
157        <br />
158        <div id="img_intro_examples" class="img_container">
159          <center><img src="./figs/table1.png" border="0" width="55%"></center>
160          <br>
161        <p style="text-align:justify">
162          Table 2: Evaluation results on LFW, CALFW, CPLFW and AgeDB: Comparison with FPVT, CNNs, PureViTs and 
163          Convolutional ViTs methods. The top first three models represents the results of most representative 
164          CNN models, while the mid block indicates the performance of pure ViTs architectures. The third
165          block shows the results of Conv. ViTs models
166        </p>
167        </div>
168     </div>
169        
170      <!--=================Citation==========================-->
171      <div class="section Citation" , id="Citation">
172        <h2>Citation</h2>
173        <div class="section bibtex">
174<pre>@InProceedings{Khawar_BMVC22_FPVT,
175  author = {Khawar Islam, Muhammad Zaigham Zaheer, Arif Mahmood},
176  title = {Face Pyramid Vision Transformer},
177  booktitle = {Proceedings of the British Machine Vision Conference},
178  year = {2022}
179}</pre>
180<pre>@inproceedings{islam2021face,
181  title={Face Recognition Using Shallow Age-Invariant Data},
182  author={Islam, Khawar and Lee, Sujin and Han, Dongil and Moon, Hyeonjoon},
183  booktitle={2021 36th International Conference on Image and Vision Computing New Zealand (IVCNZ)},
184  pages={1--6},
185  year={2021},
186  organization={IEEE}
187}</pre>
188        </div>
189     </div>
190
191      <!--=================Contact==========================-->
192      <div class="section contact", id="Contact">
193        <h2 id="contact">Contact</h2>
194        <p style="text-align:justify">If you have any question, please contact to Khawar Islam at
195          <strong>khawarr dot islam at gmail dot com</strong>.</p>
196      </div>
197</body>
198
199</html>