1 2<!DOCTYPE html> 3<html> 4 5<head lang="en"> 6 <meta charset="UTF-8"> 7 <meta http-equiv="x-ua-compatible" content="ie=edge"> 8 9 <title>StyleAdv</title> 10 11 <meta name="description" content=""> 12 <meta name="viewport" content="width=device-width, initial-scale=1"> 13 14 <!-- <base href="/"> --> 15 16 <!--FACEBOOK--> 17 <meta property="og:image" content="https://jonbarron.info/mipnerf/img/rays_square.png"> 18 <meta property="og:image:type" content="image/png"> 19 <meta property="og:image:width" content="682"> 20 <meta property="og:image:height" content="682"> 21 <meta property="og:type" content="website" /> 22 <meta property="og:url" content="https://jonbarron.info/mipnerf/"/> 23 <meta property="og:title" content="mip-NeRF" /> 24 <meta property="og:description" content="Project page for Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields." /> 25 26 <!--TWITTER--> 27 <meta name="twitter:card" content="summary_large_image" /> 28 <meta name="twitter:title" content="mip-NeRF" /> 29 <meta name="twitter:description" content="Project page for Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields." /> 30 <meta name="twitter:image" content="https://jonbarron.info/mipnerf/img/rays_square.png" /> 31 32 33<!-- <link rel="apple-touch-icon" href="apple-touch-icon.png"> --> 34 <!-- <link rel="icon" type="image/png" href="img/seal_icon.png"> --> 35 <!-- Place favicon.ico in the root directory --> 36 37 <link rel="stylesheet" href="https://maxcdn.bootstrapcdn.com/bootstrap/3.3.5/css/bootstrap.min.css"> 38 <link rel="stylesheet" href="https://maxcdn.bootstrapcdn.com/font-awesome/4.4.0/css/font-awesome.min.css"> 39 <link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/codemirror/5.8.0/codemirror.min.css"> 40 <link rel="stylesheet" href="css/app.css"> 41 42 <link rel="stylesheet" href="css/bootstrap.min.css"> 43 44
44<script src="https://ajax.googleapis.com/ajax/libs/jquery/1.11.3/jquery.min.js"></script>
44 45
45<script src="https://maxcdn.bootstrapcdn.com/bootstrap/3.3.5/js/bootstrap.min.js"></script>
45 46
46<script src="https://cdnjs.cloudflare.com/ajax/libs/codemirror/5.8.0/codemirror.min.js"></script>
46 47
47<script src="https://cdnjs.cloudflare.com/ajax/libs/clipboard.js/1.5.3/clipboard.min.js"></script>
47 48 49
49<script src="js/app.js"></script>
49 50</head> 51 52<body> 53 <div class="container" id="main"> 54 <div class="row"> 55 <h2 class="col-md-12 text-center"> 56 <b>StyleAdv: Meta Style Adversarial Training for Cross-Domain Few-Shot Learning</b> </br> 57 <small> 58 CVPR 2023 59 </small> 60 </h2> 61 </div> 62 <div class="row"> 63 <div class="col-md-12 text-center"> 64 <ul class="list-inline"> 65 <li> 66 <a href="http://yuqianfu.com/"> 67 Yuqian Fu 68 </a> 69 </br>Fudan 70 </li> 71 <li> 72 <a href=""> 73 Yu Xie 74 </a> 75 </br>Fudan 76 </li> 77 <li> 78 <a href="https://yanweifu.github.io/"> 79 Yanwei Fu 80 </a> 81 </br>Fudan 82 </li> 83 <li> 84 <a href="https://fvl.fudan.edu.cn/people/yugangjiang"> 85 Yu-Gang Jiang 86 </a> 87 </br>Fudan 88 </li> 89 </ul> 90 </div> 91 </div> 92 93 94 <div class="row"> 95 <div class="col-md-4 col-md-offset-4 text-center"> 96 <ul class="nav nav-pills nav-justified"> 97 <li> 98 <a href="https://arxiv.org/pdf/2302.09309"> 99 <image src="img/paper_icon.png" height="60px"> 100 <h4><strong>Paper</strong></h4> 101 </a> 102 </li> 103 <li> 104 <a href="https://www.bilibili.com/video/BV1th4y1s78H/?spm_id_from=333.999.0.0&vd_source=668a0bb77d7d7b855bde68ecea1232e7"> 105 <image src="img/bilibili_icon.png" height="60px"> 106 <h4><strong>Video</strong></h4> 107 </a> 108 </li> 109 <li> 110 <a href="https://youtu.be/YB-S2YF22mc"> 111 <image src="img/youtube_icon.png" height="60px"> 112 <h4><strong>Video</strong></h4> 113 </a> 114 </li> 115 <li> 116 <a href="https://github.com/lovelyqian/StyleAdv-CDFSL"> 117 <image src="img/github.png" height="60px"> 118 <h4><strong>Code</strong></h4> 119 </a> 120 </li> 121 </ul> 122 </div> 123 </div> 124 125 126 127 <div class="row"> 128 <div class="col-md-8 col-md-offset-2"> 129 <h3> 130 <b>Abstract</b> 131 </h3> 132 <image src="img/abstract.png" class="img-responsive" alt="overview"><br> 133 <p class="text-justify"> 134 Cross-Domain Few-Shot Learning (CD-FSL) is a recently emerging task that tackles few-shot learning across different domains. It aims at transferring prior knowledge learned on the source dataset to novel target datasets. The CD-FSL task is especially challenged by the huge domain gap between different datasets. Critically, such a domain gap actually comes from the changes of visual styles, and wave-SAN empirically shows that spanning the style distribution of the source data helps alleviate this issue. 135 However, wave-SAN simply swaps styles of two images. Such a vanilla operation makes the generated styles ``real'' and ``easy'', which still fall into the original set of the source styles. 136 Thus, inspired by vanilla adversarial learning, a novel model-agnostic meta Style Adversarial training (StyleAdv) method together with a novel style adversarial attack method is proposed for CD-FSL. 137 Particularly, our style attack method synthesizes both ``virtual'' and ``hard'' adversarial styles for model training. This is achieved by perturbing the original style with the signed style gradients. 138 By continually attacking styles and forcing the model to recognize these challenging adversarial styles, our model is gradually robust to the visual styles, thus boosting the generalization ability for novel target datasets. 139 Besides the typical CNN-based backbone, we also employ our StyleAdv method on large-scale pretrained vision transformer. Extensive experiments conducted on eight various target datasets show the effectiveness of our method. Whether built upon ResNet or ViT, we achieve the new state of the art for CD-FSL. 140 </p> 141 </div> 142 </div> 143 144 145 146 <div class="row"> 147 <div class="col-md-8 col-md-offset-2"> 148 <h3> 149 <b>StyleAdv Method</b> 150 </h3>
151 <p class="text-justify"> 152 StyleAdv solves CD-FSL via generating both "virtual" and "hard" styles. Generally, we have two iterative loops: 153 <ul> 154 <li>Inner loop: synthesizing new adversarial styles by attacking the original source styles; </li> 155 <li>Outer loop: optimizing the whole network by classifying source images with both original and adversarial styles; </li> 156 </ul> 157 By interatively perform the inner loop and the outer loop, the generalization ability on various styles is improved. 158 </p> 159 <p style="text-align:center;"> 160 <image src="img/styleAdv-framework.png" class="img-responsive" alt="scales"> 161 </p> 162 <p class="text-justify"> 163 The details of how we achieve the style attack and apply StyleAdv on both ResNet and ViT backbones please see the paper. 164 </p> 165 </div> 166 </div> 167 168 169 170 <div class="row"> 171 <div class="col-md-8 col-md-offset-2"> 172 <h3> 173 <b>Results</b> 174 </h3> 175 <p class="text-justify"> 176 We conduct experiments on both RN10/ViT-small with eight target targets including ChestX, ISIC, EuroSAT, CropDisease, CUB, Cars, Places, and Plantae. Both the results of StyleAdv (only meta-trained) and StyleAdv-FT (finetuned with few target support images) are reported. 177 </p> 178 <p style="text-align:center;"> 179 <image src="img/styleAdv-result.png" class="img-responsive" alt="scales"> 180 </p> 181 <p class="text-justify"> 182 We also provide the visulization result. Note that wave-SAN is our former work for StyleAdv which also augments style but in a easier way. 183 </p> 184 <p style="text-align:center;"> 185 <image src="img/styleAdv-vis.png" class="img-responsive" alt="scales"> 186 </p> 187 </div> 188 </div> 189 190 191 <div class="row"> 192 <div class="col-md-8 col-md-offset-2"> 193 <h3> 194 <b>Related Works</b> 195 </h3> 196 <p class="text-justify"> 197 <a href="https://github.com/lovelyqian/wave-SAN-CDFSL">Wave-SAN</a> is the former work that tackles CD-FSL via style augmentation. 198 </p> 199 <p class="text-justify"> 200 <a href="https://github.com/lovelyqian/Meta-FDMixup">Meta-FDMixup</a>, <a href="https://github.com/lovelyqian/ME-D2N_for_CDFSL">Me-D2N</a>, and <a href="https://arxiv.org/abs/2210.05392">TGDM</a> are our works that address CD-FSL with few-labeled target examples. 201 </p> 202 </div> 203 </div> 204 205 206 <div class="row"> 207 <div class="col-md-8 col-md-offset-2"> 208 <h3> 209 <b>Citation</b> 210 </h3> 211 <div class="form-group col-md-10 col-md-offset-1"> 212 <textarea id="bibtex" class="form-control" readonly> 213@inproceedings{fu2023styleadv, 214 title={StyleAdv: Meta Style Adversarial Training for Cross-Domain Few-Shot 215 Learning}, 216 author={Fu, Yuqian and Xie, Yu and Fu, Yanwei and Jiang, Yu-Gang}, 217 booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and 218 Pattern Recognition}, 219 pages={24575--24584}, 220 year={2023} 221} 222</textarea> 223 </div> 224 </div> 225 </div> 226 </div> 227</body> 228</html>
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.