1<!DOCTYPE html> 2<html lang="en"> 3<head> 4 <meta http-equiv="Content-Type" content="text/html; charset=UTF-8"/> 5 <meta name="viewport" content="width=device-width, initial-scale=1, maximum-scale=1.0"/> 6 <meta name="author" content="jasonlai"> 7 <title>Every Pixel Matters: Center-aware Feature Alignment for Domain Adaptive Object Detector</title> 8 9 <!-- CSS --> 10 <link href="https://fonts.googleapis.com/icon?family=Material+Icons" rel="stylesheet"> 11 <link href="css/materialize.css" type="text/css" rel="stylesheet" media="screen,projection"/> 12 <link href="css/style.css" type="text/css" rel="stylesheet" media="screen,projection"/> 13 <link href="css/font-awesome.min.css" rel="stylesheet"> 14 15 <!--<meta property="og:image" content="http://gph.is/2oZQz8h" />--> 16 17 <!-- Latex support --> 18
18<script type="text/x-mathjax-config"> MathJax.Hub.Config({ TeX: { equationNumbers: { autoNumber: "all" } } }); </script>
18 19
19<script type="text/x-mathjax-config"> 20 MathJax.Hub.Config({ 21 tex2jax: { 22 inlineMath: [ ['$','$'], ["\\(","\\)"] ], 23 processEscapes: true 24 } 25 }); 26 </script>
26 27
27<script src='https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.4/latest.js?config=TeX-MML-AM_CHTML' async></script>
27 28</head> 29 30<body> 31 32 <div class="navbar-fixed"> 33 34 <nav class="grey darken-4" role="navigation"> 35 <div class="nav-wrapper container"><a id="logo-container" href="#" class="brand-logo"></a> 36 <ul class="left hide-on-med-and-down"> 37 <li><a class="nav-item waves-effect waves-light" href="#home">Home</a></li> 38 <li><a class="nav-item waves-effect waves-light" href="#abstract">Abstract</a></li> 39 <li><a class="nav-item waves-effect waves-light" href="#paper">Paper</a></li> 40 <li><a class="nav-item waves-effect waves-light" href="#citation">Citation</a></li> 41 <li><a class="nav-item waves-effect waves-light" href="#download">Results</a></li> 42 <li><a class="nav-item waves-effect waves-light" href="#reference">References</a></li> 43 </ul> 44 <a href="#" data-activates="nav-mobile" class="button-collapse"><i class="material-icons">menu</i></a> 45 </div> 46 </nav> 47 </div> 48 49 50 <div class="section no-pad-bot" id="index-banner"> 51 <div class="container scrollspy" id="home"> 52 53 <h4 class="header center black-text">Every Pixel Matters: Center-aware Feature Alignment for <br>Domain Adaptive Object Detector</h4> 54 55 <br> 56 57 <div class="row center"> 58 <h5 class="header col m3 s12"> 59 <div class="author"><a href = "https://chengchunhsu.github.io/" target="blank">Cheng-Chun Hsu</a></div> 60 <div class="school"><a href= "https://www.sinica.edu.tw/en" target="blank">Academia Sinica</a></div> 61 </h5> 62 63 <h5 class="header col m3 s12"> 64 <div class="author"><a href="https://sites.google.com/site/yihsuantsai/" target="blank">Yi-Hsuan Tsai</a></div> 65 <div class="school"><a href="http://www.nec-labs.com/research-departments/media-analytics/media-analytics-home" target="blank">NEC Labs America</a></div> 66 </h5> 67 68 <h5 class="header col m3 s12"> 69 <div class="author"><a href="https://sites.google.com/site/yylinweb/" target="blank">Yen-Yu Lin</a></div> 70 <div class="school"><a href="https://www.sinica.edu.tw/en" target="blank">Academia Sinica</a></div> 71 <div class="school"><a href="https://www.nctu.edu.tw/en" target="blank">National Chiao Tung University</a></div> 72 </h5> 73 74 <h5 class="header col m3 s12"> 75 <div class="author"><a href="https://faculty.ucmerced.edu/mhyang/" target="blank">Ming-Hsuan Yang</a></div> 76 <div class="school"><a href="https://www.ucmerced.edu/" target="blank">UC Merced</a></div> 77 <div class="school"><a href="https://research.google/" target="blank">Google Research</a></div> 78 </h5> 79 80 </div> 81 82 </div> 83 </div> 84 85 <div class="container"> 86 87 <div class="section"> 88 89 <!-- Icon Section --> 90 <div class="row center"> 91 <div class="col l12 m12 s12"> 92 <img class="responsive-img" src="img/teaser.png" width="70%"> 93 </a> 94 </div> 95 </div> 96 97 </div> 98 99 <div class="row section scrollspy" id="abstract"> 100 <div class="title">Abstract</div> 101 <br> 102 <p align="justify"> 103 A domain adaptive object detector aims to adapt itself to unseen domains that may contain variations of object 104 appearance, viewpoints or backgrounds. 105 Most existing methods adopt feature alignment either on the image level or instance level. 106 However, image-level alignment on global features may tangle foreground/background pixels at the same time, while 107 instance-level alignment using proposals may suffer from the background noise. 108 Different from existing solutions, we propose a domain adaptation framework that accounts for each pixel via predicting 109 pixel-wise objectness and centerness. 110 Specifically, the proposed method carries out center-aware alignment by paying more attention to foreground pixels, 111 hence achieving better adaptation across domains. 112 We demonstrate our method on numerous adaptation settings with extensive experimental results and show favorable 113 performance against existing state-of-the-art algorithms. 114 </p> 115 </div> 116 117 <div class="section"> 118 <!-- Icon Section --> 119 <div class="row center"> 120 <div class="col l12 m12 s12"> 121 <img class="responsive-img" src="img/architecture.png" width="70%">
122 <figcaption> 123 <p align="justify"> 124 <b>Fig. Proposed framework for domain adaptive object detection.</b> 125 Given the source and target images, we feed them to a shared feature extractor $G$ to obtain their features $F$. 126 Then, the global alignment on these features is performed via a global discriminator $D_{GA}$ and a domain prediction loss $L_{GA}$. 127 Next, we pass the feature through the fully-convolutional module $P$ to produce the classi�cation and centerness maps. 128 These maps and the feature $F$ are utilized to generate the center-aware features. 129 Finally, we use a center-aware discriminator $D_{CA}$ and another domain prediction loss $L_{CA}$ to perform the proposed center-aware feature alignment. 130 Note that the bounding box prediction loss $L_{det}$ is only operated on source images using their corresponding ground-truth bounding boxes. 131 </p> 132 </figcaption> 133 </a> 134 </div> 135 </div> 136 </div> 137 138 <div class="row section scrollspy" id="paper"> 139 <div class="title">Papers</div> 140 <br> 141 <div class="row"> 142 143 <div class="col m2 s6 center"> 144 <a href="https://arxiv.org/pdf/2008.08574" target=""> 145 <img src="img/icon_pdf.png" width="60%"> 146 </a> 147 <br> 148 <a href="" target="">ECCV 2020<br></a> 149 </div> 150 151 </div> 152 153 <div class="row section scrollspy" id="citation"> 154 <div class="title">Citation</div> 155 <p>Cheng-Chun Hsu, Yi-Hsuan Tsai, Yen-Yu Lin, and Ming-Hsuan Yang, "Every Pixel Matters: Center-aware Feature Alignment for Domain Adaptive Object Detector", in European Conference on Computer Vision (ECCV), 2020.</p> 156 157 <br> 158 159 <div class="title">BibTex</div> 160 <pre> 161@inproceedings{hsu2020epm, 162 title = {Every Pixel Matters: Center-aware Feature Alignment for Domain Adaptive Object Detector}, 163 author = {Cheng-Chun Hsu, Yi-Hsuan Tsai, Yen-Yu Lin, Ming-Hsuan Yang}, 164 booktitle = {European Conference on Computer Vision}, 165 year = {2020} 166} 167 </pre> 168 169 </div> 170 171 <div class="section row scrollspy" id="download"> 172 <div class="title">Code and Results</div> 173 <br> 174 <div class="row"> 175 176 <div class="col m6 s12 center"> 177 <a href="https://github.com/chengchunhsu/EveryPixelMatters" target=""> 178 <img src="img/github.png"> 179 </a> 180 <br> 181 <a href="https://github.com/chengchunhsu/EveryPixelMatters" target="">Code</a> 182 </div> 183 184 <div class="col m6 s12 center"> 185 <a href="https://github.com/chengchunhsu/EveryPixelMatters" target=""> 186 <img src="img/icon_zip.png"> 187 </a> 188 <br> 189 <a href="https://github.com/chengchunhsu/EveryPixelMatters" >Results</a> 190 </div> 191 192 </div> 193 </div> 194 195 <div class="row section scrollspy" id="reference"> 196 <div class="title">References</div> 197 <ul> 198 <li>• 199 Chen et al. <a 200 href="https://openaccess.thecvf.com/content_cvpr_2018/html/Chen_Domain_Adaptive_Faster_CVPR_2018_paper.html" 201 target="blank">Domain Adaptive Faster R-CNN for Object Detection in the Wild.</a> 202 In CVPR, 2018. 203 </li> 204 205 <li>• 206 Saito et al. <a 207 href="https://openaccess.thecvf.com/content_CVPR_2019/html/Saito_Strong-Weak_Distribution_Alignment_for_Adaptive_Object_Detection_CVPR_2019_paper.html" 208 target="blank">Strong-weak distribution alignment for adaptive object detection.</a> 209 In CVPR, 2019. 210 </li> 211 212 <li>• 213 Zhu et al. <a 214 href="https://openaccess.thecvf.com/content_CVPR_2019/papers/Zhu_Adapting_Object_Detectors_via_Selective_Cross-Domain_Alignment_CVPR_2019_paper.pdf" 215 target="blank">Adapting object detectors via selective cross-domain alignment.</a> 216 In CVPR, 2019. 217 </li> 218 219 <li>• 220 He et al. <a 221 href="https://openaccess.thecvf.com/content_ICCV_2019/html/He_Multi-Adversarial_Faster-RCNN_for_Unrestricted_Object_Detection_ICCV_2019_paper.html" 222 target="blank">
222Multi-adversarial faster-rcnn for unrestricted object detection.</a> 223 In ICCV, 2019. 224 </li> 225 226 <li>• 227 Kim et al. <a 228 href="https://openaccess.thecvf.com/content_CVPR_2019/html/Kim_Diversify_and_Match_A_Domain_Adaptive_Representation_Learning_Paradigm_for_CVPR_2019_paper.html" 229 target="blank">Diversify and match: A domain adaptive representation learning paradigm for object detection.</a> 230 In CVPR, 2019. 231 </li> 232 233 <li>• 234 Cai et al. <a 235 href="https://openaccess.thecvf.com/content_CVPR_2019/html/Cai_Exploring_Object_Relation_in_Mean_Teacher_for_Cross-Domain_Detection_CVPR_2019_paper.html" 236 target="blank">Exploring object relation in mean teacher for cross-domain detection.</a> 237 In CVPR, 2019. 238 </li> 239 240 <li>• 241 Hsu et al. <a 242 href="https://arxiv.org/abs/1910.11319" 243 target="blank">Progressive domain adaptation for object detection.</a> 244 In WACV, 2020. 245 </li> 246 247 <li>• 248 Kim et al. <a 249 href="https://openaccess.thecvf.com/content_ICCV_2019/html/Kim_Self-Training_and_Adversarial_Background_Regularization_for_Unsupervised_Domain_Adaptive_One-Stage_ICCV_2019_paper.html" 250 target="blank">Self-training and adversarial background regularization for unsupervised domain adaptive one-stage object detection.</a> 251 In ICCV, 2019. 252 </li> 253 254 <li>• 255 Ren et al. <a 256 href="https://papers.nips.cc/paper/5638-faster-r-cnn-towards-real-time-object-detection-with-region-proposal-networks" 257 target="blank">Faster r-cnn: Towards real-time object detection with region proposal networks.</a> 258 In NeurIPS, 2015. 259 </li> 260 261 <li>• 262 Tian et al. <a 263 href="https://openaccess.thecvf.com/content_ICCV_2019/html/Tian_FCOS_Fully_Convolutional_One-Stage_Object_Detection_ICCV_2019_paper.html" 264 target="blank">Fcos: Fully convolutional one-stage object detection.</a> 265 In ICCV, 2019. 266 </li> 267 268 </ul> 269 </div> 270 </div> 271 272 273 <!-- Scripts--> 274
274<script src="https://code.jquery.com/jquery-2.1.1.min.js"></script>
274 275
275<script src="js/materialize.js"></script>
275 276
276<script src="js/init.js"></script>
276 277 278 </body> 279</html> 280 281
Line numbers count LF bytes from the start of the resource, as the search results do. Vendor segments are library code the classifier recognised; they are stored but not indexed. Bytes are shown as Latin1 characters, one per byte.