<html xmlns="http://www.w3.org/1999/xhtml"><head><meta charset="utf-8"><meta name="generator" content="pdf2htmlEX"><meta http-equiv="X-UA-Compatible" content="IE=edge,chrome=1"><link rel="stylesheet" href="https://csdnimg.cn/release/download_crawler_static/css/base.min.css"><link rel="stylesheet" href="https://csdnimg.cn/release/download_crawler_static/css/fancy.min.css"><link rel="stylesheet" href="https://csdnimg.cn/release/download_crawler_static/10882440/raw.css"><script src="https://csdnimg.cn/release/download_crawler_static/js/compatibility.min.js"></script><script src="https://csdnimg.cn/release/download_crawler_static/js/pdf2htmlEX.min.js"></script><script>try{pdf2htmlEX.defaultViewer = new pdf2htmlEX.Viewer({});}catch(e){}</script><title></title></head><body><div id="sidebar" style="display: none"><div id="outline"></div></div><div id="pf1" class="pf w0 h0" data-page-no="1"><div class="pc pc1 w0 h0"><img class="bi x0 y0 w1 h1" alt="" src="https://csdnimg.cn/release/download_crawler_static/10882440/bg1.jpg"><div class="t m0 x1 h2 y1 ff1 fs0 fc0 sc0 ls0 ws0"><span class="fc1 sc0">CV</span><span class="fc1 sc0">PR</span><span class="fc1 sc0">-</span><span class="ls1"><span class="fc1 sc0">2018</span></span><span class="fc1 sc0"> </span><span class="_ _0"> </span><span class="ff2"><span class="fc1 sc0">阿里</span><span class="fc1 sc0">巴巴</span><span class="fc1 sc0">收录</span><span class="fc1 sc0">论文</span></span><span class="fc1 sc0"> </span></div><div class="t m0 x2 h3 y2 ff1 fs0 fc0 sc0 ls0 ws0"><span class="fc1 sc0"> </span></div><div class="t m0 x3 h4 y3 ff3 fs1 fc0 sc0 ls0 ws0"><span class="fc1 sc0"> </span></div><div class="t m0 x3 h4 y4 ff3 fs1 fc0 sc0 ls0 ws0"><span class="fc1 sc0"> </span></div><div class="t m0 x3 h4 y5 ff3 fs1 fc0 sc0 ls0 ws0"><span class="fc1 sc0"> </span></div><div class="t m0 x3 h4 y6 ff3 fs1 fc0 sc0 ls0 ws0"><span class="fc1 sc0"> </span></div><div class="t m0 x3 h4 y7 ff3 fs1 fc0 sc0 ls0 ws0"><span class="fc1 sc0"> </span></div><div class="t m0 x3 h4 y8 ff3 fs1 fc0 sc0 ls0 ws0"><span class="fc1 sc0"> </span></div><div class="t m0 x3 h4 y9 ff3 fs1 fc0 sc0 ls0 ws0"><span class="fc1 sc0"> </span></div><div class="t m0 x3 h4 ya ff3 fs1 fc0 sc0 ls0 ws0"><span class="fc1 sc0"> </span></div><div class="t m0 x3 h4 yb ff3 fs1 fc0 sc0 ls0 ws0"><span class="fc1 sc0"> </span></div><div class="t m0 x3 h4 yc ff3 fs1 fc0 sc0 ls0 ws0"><span class="fc1 sc0"> </span></div><div class="t m0 x3 h4 yd ff3 fs1 fc0 sc0 ls0 ws0"><span class="fc1 sc0"> </span></div><div class="t m0 x3 h4 ye ff3 fs1 fc0 sc0 ls0 ws0"><span class="fc1 sc0"> </span></div><div class="t m0 x3 h4 yf ff3 fs1 fc0 sc0 ls0 ws0"><span class="fc1 sc0"> </span></div><div class="t m0 x3 h4 y10 ff3 fs1 fc0 sc0 ls0 ws0"><span class="fc1 sc0"> </span></div><div class="t m0 x3 h4 y11 ff3 fs1 fc0 sc0 ls0 ws0"><span class="fc1 sc0"> </span></div><div class="t m0 x3 h4 y12 ff3 fs1 fc0 sc0 ls0 ws0"><span class="fc1 sc0"> </span></div><div class="t m0 x3 h4 y13 ff3 fs1 fc0 sc0 ls0 ws0"><span class="fc1 sc0"> </span></div><div class="t m0 x3 h4 y14 ff3 fs1 fc0 sc0 ls0 ws0"><span class="fc1 sc0"> </span></div><div class="t m0 x3 h4 y15 ff3 fs1 fc0 sc0 ls0 ws0"><span class="fc1 sc0"> </span></div><div class="t m0 x3 h4 y16 ff3 fs1 fc0 sc0 ls0 ws0"><span class="fc1 sc0"> </span></div><div class="t m0 x3 h4 y17 ff3 fs1 fc0 sc0 ls0 ws0"><span class="fc1 sc0"> </span></div><div class="t m0 x3 h4 y18 ff3 fs1 fc0 sc0 ls0 ws0"><span class="fc1 sc0"> </span></div><div class="t m0 x3 h4 y19 ff3 fs1 fc0 sc0 ls0 ws0"><span class="fc1 sc0"> </span></div></div><div class="pi" data-data='{"ctm":[1.611792,0.000000,0.000000,1.611792,0.000000,0.000000]}'></div></div></body></html>
<div id="pf2" class="pf w0 h0" data-page-no="2"><div class="pc pc2 w0 h0"><img class="bi x0 y0 w1 h1" alt="" src="https://csdnimg.cn/release/download_crawler_static/10882440/bg2.jpg"><div class="t m0 x1 h2 y1 ff1 fs0 fc0 sc0 ls0 ws0">CVPR-<span class="ls1">2018</span> <span class="_ _0"> </span><span class="ff2">阿里巴巴收录论文</span> </div><div class="t m0 x2 h3 y2 ff1 fs0 fc0 sc0 ls0 ws0"> </div><div class="t m0 x2 h4 y3 ff3 fs1 fc0 sc0 ls0 ws0"> </div><div class="t m0 x3 h4 y4 ff3 fs1 fc0 sc0 ls0 ws0"> </div><div class="t m0 x3 h4 y5 ff3 fs1 fc0 sc0 ls0 ws0"> </div><div class="t m0 x4 h5 y1a ff4 fs2 fc0 sc0 ls0 ws0"> </div><div class="t m0 x5 h6 y1b ff4 fs3 fc0 sc0 ls0 ws0"> </div><div class="t m0 x4 h7 y1c ff4 fs4 fc0 sc0 ls0 ws0"> </div><div class="t m0 x6 h8 y1d ff3 fs5 fc2 sc0 ls0 ws0"> </div><div class="t m0 x7 h8 y1e ff3 fs5 fc2 sc0 ls0 ws0"> </div><div class="t m0 x7 h8 y1f ff2 fs5 fc2 sc0 ls0 ws0">扫一扫二维码图案，关<span class="_ _1"></span>注我吧<span class="ff3"> </span></div><div class="t m0 x4 h9 y20 ff4 fs6 fc0 sc0 ls0 ws0"> </div><div class="t m0 x8 ha y21 ff4 fs5 fc0 sc0 ls0 ws0"><span class="fc1 sc0"> </span></div><div class="t m0 x8 ha y22 ff4 fs5 fc0 sc0 ls0 ws0"><span class="fc1 sc0"> </span></div><div class="t m0 x8 ha y23 ff4 fs5 fc0 sc0 ls0 ws0"><span class="fc1 sc0"> </span></div><div class="t m0 x9 hb y24 ff2 fs7 fc2 sc0 ls0 ws0">「阿里技术」微信公众<span class="_ _2"></span>号<span class="ff3"> <span class="_ _3"> </span></span>「阿里巴巴机器智能」<span class="_ _2"></span>微信公众号<span class="ff3 fc0"> </span></div><div class="t m0 x2 h4 y25 ff3 fs1 fc0 sc0 ls0 ws0"> </div><div class="t m0 x3 h4 y26 ff3 fs1 fc0 sc0 ls0 ws0"> </div><div class="t m0 x3 h4 y27 ff3 fs1 fc0 sc0 ls0 ws0"> </div><div class="t m0 x3 h4 y28 ff3 fs1 fc0 sc0 ls0 ws0"> </div><div class="t m0 x3 h4 y29 ff3 fs1 fc0 sc0 ls0 ws0"> </div><div class="t m0 x3 h4 y2a ff3 fs1 fc0 sc0 ls0 ws0"> </div><div class="t m0 x3 h4 y2b ff3 fs1 fc0 sc0 ls0 ws0"> </div><div class="t m0 x3 h4 y2c ff3 fs1 fc0 sc0 ls0 ws0"> </div><div class="t m0 xa h4 y2d ff2 fs1 fc3 sc0 ls0 ws0">本书著作权归阿<span class="_ _2"></span>里巴巴集团所有<span class="_ _2"></span>，<span class="ff3"> </span></div><div class="t m0 xb h4 y2e ff2 fs1 fc3 sc0 ls0 ws0">未经授权不得进<span class="_ _2"></span>行转载或其他任<span class="_ _2"></span>何形式的二次<span class="_ _2"></span>传播。<span class="ff3 fc0"> </span></div><div class="t m0 x2 h4 y2f ff3 fs1 fc0 sc0 ls0 ws0"> </div></div><div class="pi" data-data='{"ctm":[1.611792,0.000000,0.000000,1.611792,0.000000,0.000000]}'></div></div>
<div id="pf3" class="pf w0 h0" data-page-no="3"><div class="pc pc3 w0 h0"><img class="bi x0 y0 w1 h1" alt="" src="https://csdnimg.cn/release/download_crawler_static/10882440/bg3.jpg"><div class="t m0 x1 h2 y1 ff1 fs0 fc0 sc0 ls0 ws0">CVPR-<span class="ls1">2018</span> <span class="_ _0"> </span><span class="ff2">阿里巴巴收录论文</span> </div><div class="t m0 x2 h3 y2 ff1 fs0 fc0 sc0 ls0 ws0"> </div><div class="t m0 x3 h4 y30 ff3 fs1 fc0 sc1 ls0 ws0">| <span class="ff2">序言</span> </div><div class="t m0 x3 h4 y31 ff3 fs1 fc0 sc0 ls0 ws0">CVPR<span class="ff2">（</span>Conferen<span class="_ _2"></span>ce on Com<span class="_ _2"></span>puter Vision a<span class="_ _2"></span>nd Patter<span class="_ _2"></span>n Recognition<span class="_ _2"></span><span class="ff2">）是计算机<span class="_ _2"></span>视觉领域的顶</span></div><div class="t m0 x2 h4 y32 ff2 fs1 fc0 sc0 ls0 ws0">会之一，<span class="ls2">伴随</span>着<span class="ls2">视觉<span class="_ _0"> </span></span><span class="ff3">AI<span class="_ _0"> </span></span>的<span class="_ _2"></span>火热<span class="ls2">，近</span>几年参会人数急剧增加<span class="_ _2"></span>。<span class="ff3">2018<span class="_ _0"> </span></span>年的<span class="_ _0"> </span><span class="ff3">C<span class="_ _2"></span>VPR<span class="_ _0"> </span><span class="ff2">会议于<span class="_ _0"> </span></span>6<span class="_ _4"> </span><span class="ff2">月<span class="_ _0"> </span></span>18<span class="_ _4"> </span><span class="ff2">日</span>-22</span></div><div class="t m0 x2 h4 y33 ff2 fs1 fc0 sc0 ls0 ws0">日在美国犹他州<span class="_ _2"></span>盐湖城举办。本<span class="_ _2"></span>届大会有超过<span class="_ _2"></span><span class="ff3"> <span class="_"> </span>3300 <span class="_"> </span><span class="ff2">篇<span class="_ _2"></span>的大会论文投<span class="_ _2"></span>稿，录取<span class="ff3"> <span class="_"> </span>97<span class="_ _2"></span>9 <span class="_"> </span><span class="ff2">篇（接受</span></span></span></span></div><div class="t m0 x2 h4 y34 ff2 fs1 fc0 sc0 ls0 ws0">率约为<span class="ff3"> <span class="_"> </span>29%</span>）<span class="_ _2"></span>，其中包括<span class="_ _2"></span><span class="ff3"> <span class="_"> </span>70 <span class="_"> </span><span class="ff2">篇</span> <span class="_"> </span>Oral<span class="_ _2"></span> <span class="_"> </span><span class="ff2 ls2">论文<span class="ls0">和</span></span> <span class="_"> </span>224 <span class="_"> </span><span class="ff2">篇</span> <span class="_"> </span>Spo<span class="_ _2"></span>tlight <span class="_"> </span><span class="ff2">论文，<span class="_ _2"></span>参会人数达到</span></span></div><div class="t m0 x2 h4 y35 ff3 fs1 fc0 sc0 ls0 ws0">6500<span class="_ _4"> </span><span class="ff2">人。除了正会以外<span class="_ _2"></span>，本届<span class="_ _0"> </span><span class="ff3">CVPR<span class="_ _4"> </span></span>有<span class="_ _0"> </span><span class="ff3">21<span class="_ _0"> </span></span>个<span class="_ _0"> </span><span class="ff3">tutor<span class="_ _2"></span>ials<span class="_ _0"> </span><span class="ff2">和<span class="_ _4"> </span></span>48<span class="_ _0"> </span><span class="ff2">个<span class="_ _0"> </span></span>workshops<span class="ff2">，<span class="_ _2"></span>以及<span class="ls2">超过<span class="_ _0"> </span></span><span class="ff3">115<span class="_ _0"> </span></span>个公司</span></span></span></div><div class="t m0 x2 h4 y36 ff2 fs1 fc0 sc0 ls0 ws0">的工业展会。<span class="_ _2"></span><span class="ff3"> </span></div><div class="t m0 x3 h4 y37 ff2 fs1 fc0 sc0 ls0 ws0">近些年伴随着深<span class="_ _2"></span>度学习技术、<span class="ff3">GP<span class="_ _2"></span>U<span class="_ _0"> </span><span class="ff2">和云计算等<span class="_ _2"></span>运算力的增强，<span class="_ _2"></span>计算机视觉技<span class="_ _2"></span>术进入越<span class="ls2">来越实用</span></span></span></div><div class="t m0 x2 h4 y38 ff2 fs1 fc0 sc0 ls0 ws0">的阶段。无论是<span class="_ _2"></span>在<span class="ls2">电商</span>、安防、娱乐，还<span class="_ _2"></span>是在工业、医疗<span class="_ _2"></span>、<span class="ls2">自动</span>驾驶，计算机视觉<span class="_ _2"></span>技术都扮演着</div><div class="t m0 x2 h4 y39 ff2 fs1 fc0 sc0 ls0 ws0">越发重要的角色<span class="_ _2"></span>。在阿里巴巴广<span class="_ _2"></span>阔的商业和数<span class="_ _2"></span>据生态的发展<span class="ls2">中，</span>计算机<span class="_ _2"></span>视觉技术的研发<span class="_ _2"></span>和商业化</div><div class="t m0 x2 h4 y3a ff2 fs1 fc0 sc0 ls0 ws0">落地密不可分。<span class="_ _2"></span>比如拍立淘利用<span class="_ _2"></span>图像搜索和识<span class="_ _2"></span>别技术帮助淘宝<span class="_ _2"></span>、<span class="ls2">天猫</span>、<span class="ff3">AliExpress,<span class="_ _2"></span> Lazada<span class="_ _4"> </span><span class="ff2">等电</span></span></div><div class="t m0 x2 h4 y3b ff2 fs1 fc0 sc0 ls0 ws0">商<span class="_ _0"> </span><span class="ff3">app<span class="_ _4"> </span></span>的<span class="ls2">用户<span class="_ _1"></span></span>在移动端通<span class="_ _2"></span>过拍照就能找到<span class="_ _2"></span>相同相似的商品<span class="_ _2"></span>，从而进行更<span class="_ _2"></span>加方便的<span class="ls2">购物</span>。比如在线</div><div class="t m0 x2 h4 y3c ff2 fs1 fc0 sc0 ls0 ws0">下新零售领域，<span class="_ _2"></span>阿里研发了人的<span class="_ _2"></span>追踪和空间定<span class="_ _2"></span>位、货架商品<span class="_ _4"> </span><span class="ff3">SKU<span class="_ _0"> </span></span><span class="ls2">识别</span>等技术去推动商场<span class="_ _2"></span>、超市、</div><div class="t m0 x2 h4 y3d ff2 fs1 fc0 sc0 ls0 ws0">酒店等的人货场<span class="_ _2"></span>数字化，并在此<span class="_ _2"></span>基础上做进一<span class="_ _2"></span>步的商业分析。<span class="_ _2"></span>在城市大脑项<span class="_ _2"></span>目，<span class="ls2">阿里</span>研发了大规</div><div class="t m0 x2 h4 y3e ff2 fs1 fc0 sc0 ls0 ws0">模视频高效处理<span class="_ _2"></span>，人和车辆的<span class="ls2">搜索</span>和识别<span class="_ _2"></span>等技术<span class="ls2">帮助</span>城市交通事故<span class="_ _2"></span>识别，人流轨迹<span class="_ _2"></span>判断以及交通</div><div class="t m0 x2 h4 y3f ff2 fs1 fc0 sc0 ls0 ws0">数据样本汇总。<span class="_ _2"></span><span class="ff3"> </span></div><div class="t m0 x3 h4 y40 ff2 fs1 fc0 sc0 ls0 ws0">在本届<span class="_ _4"> </span><span class="ff3">CVPR<span class="_ _0"> </span></span><span class="ls2">顶会</span>中，阿里巴巴总<span class="_ _2"></span>共发表<span class="_ _0"> </span><span class="ff3 ls2">18<span class="_ _0"> </span></span>篇论文。此外，<span class="_ _2"></span>阿里巴巴也举<span class="_ _2"></span>办了展台展示、<span class="_ _2"></span>学</div><div class="t m0 x2 h4 y41 ff2 fs1 fc0 sc0 ls0 ws0">者晚宴、展台技<span class="_ _2"></span>术<span class="_ _0"> </span><span class="ff3">Talk<span class="_ _4"> </span></span>等多项活动，<span class="_ _2"></span>把包括图像搜索、城<span class="_ _2"></span>市大脑、自动<span class="_ _2"></span>驾驶、<span class="ff3">FashionA<span class="_ _2"></span>I<span class="ff2">、鹿班</span></span></div><div class="t m0 x2 h4 y42 ff2 fs1 fc0 sc0 ls0 ws0">设计、三维物体<span class="_ _2"></span>建模、交互仿真<span class="_ _2"></span>虚拟人、广告<span class="_ _2"></span>、多媒体智能审<span class="_ _2"></span>核等阿里巴巴<span class="_ _2"></span>在<span class="_ _0"> </span><span class="ff3">CV<span class="_ _0"> </span></span>领域的技<span class="_ _2"></span>术成果</div><div class="t m0 x2 h4 y43 ff2 fs1 fc0 sc0 ls0 ws0">和应用情况集中<span class="_ _2"></span>亮相国际舞台。<span class="_ _2"></span>在这本论文合<span class="_ _2"></span>集中，我们收录<span class="_ _2"></span>了其中有代表<span class="_ _2"></span>性的<span class="_ _0"> </span><span class="ff3">7<span class="_ _0"> </span></span>篇论文。<span class="_ _2"></span><span class="ff3"> </span></div><div class="t m0 x3 h4 y44 ff3 fs1 fc0 sc0 ls0 ws0">Spotlight<span class="_ _4"> </span><span class="ff2">论文《基于时间尺度<span class="_ _2"></span>选择的在线行<span class="_ _2"></span>为预测》讨论了<span class="_ _2"></span>视频中行为预<span class="_ _2"></span>测的一个非常重<span class="_ _2"></span>要</span></div><div class="t m0 x2 h4 y45 ff2 fs1 fc0 sc0 ls0 ws0">的问题：怎么去<span class="_ _2"></span>选择一个好的时<span class="_ _2"></span>间维度窗口？<span class="_ _2"></span>论文提出了多个<span class="_ _2"></span>子网络的尺度<span class="_ _2"></span>选择网，<span class="ls2">包括</span>时间序</div><div class="t m0 x2 h4 y46 ff2 fs1 fc0 sc0 ls0 ws0">列建模的一维卷<span class="_ _2"></span>积子网络，尺度<span class="_ _2"></span>回归子网络，<span class="_ _2"></span>以及行为预测子<span class="_ _2"></span>网络。在两个<span class="_ _2"></span>公开数据集上，<span class="_ _2"></span>尺度</div><div class="t m0 x2 h4 y47 ff2 fs1 fc0 sc0 ls0 ws0">选择网的实验结<span class="_ _2"></span>果优于其他方法<span class="_ _2"></span>，并且准确率<span class="_ _2"></span>也接近使用<span class="_ _4"> </span><span class="ff3">Ground Truth<span class="_ _4"> </span></span>尺度的结果。<span class="_ _2"></span><span class="ff3"> </span></div><div class="t m0 x3 h4 y48 ff3 fs1 fc0 sc0 ls0 ws0">Spotlight<span class="_ _4"> </span><span class="ff2">论文《基于语境对比<span class="_ _2"></span>特征和门控多<span class="_ _2"></span>尺度融合的场景<span class="_ _2"></span>分割》致力于<span class="_ _2"></span>场景分割中的两<span class="_ _2"></span>大</span></div><div class="t m0 x2 h4 y49 ff2 fs1 fc0 sc0 ls0 ws0">问题：场景图片<span class="_ _2"></span>中像素形式的多<span class="_ _2"></span>样化（例如，<span class="_ _2"></span>显著或者不显著<span class="_ _2"></span>，前景或者背<span class="_ _2"></span>景）和场景图片<span class="_ _2"></span>中物</div><div class="t m0 x2 h4 y4a ff2 fs1 fc0 sc0 ls0 ws0">体大小的多样性<span class="_ _2"></span>。文章针对这两<span class="_ _2"></span>个问题分别提<span class="_ _2"></span>出了语境对比局<span class="_ _2"></span>部特征和门控<span class="_ _2"></span>多尺度融合方法<span class="_ _2"></span>。本</div><div class="t m0 x2 h4 y4b ff2 fs1 fc0 sc0 ls0 ws0">文提出的模型<span class="_ _2"></span>在<span class="_ _0"> </span><span class="ff3">Pascal C<span class="_ _2"></span>ontext, SUN-RG<span class="_ _2"></span>BD<span class="_ _4"> </span><span class="ff2">和<span class="_ _0"> </span></span>COCO Stuff <span class="ff2">三个场<span class="_ _2"></span>景分割数据<span class="_ _2"></span>集上验证了性能<span class="_ _2"></span>，</span></span></div><div class="t m0 x2 h4 y4c ff2 fs1 fc0 sc0 ls0 ws0">取得了目前最高<span class="_ _2"></span>的场景分割性能<span class="_ _2"></span>。<span class="ff3"> </span></div><div class="t m0 x3 h4 y4d ff2 fs1 fc0 sc0 ls0 ws0">对于跨模态检索<span class="_ _2"></span>而言，如何学到<span class="_ _2"></span>合适的特征表<span class="_ _2"></span>达非常关键。<span class="ff3">Sp<span class="_ _2"></span>otlight<span class="_ _4"> </span><span class="ff2">论文《所见所想所找<span class="_ _2"></span>－</span></span></div><div class="t m0 x2 h4 y4e ff2 fs1 fc0 sc0 ls0 ws0">基于生成模型的<span class="_ _2"></span>跨模态检索》提<span class="_ _2"></span>出了一种基于<span class="_ _2"></span>生成模型的跨模<span class="_ _2"></span>态检索方法，<span class="_ _2"></span>该方法可以学习<span class="_ _2"></span>跨模</div><div class="t m0 x2 h4 y4f ff2 fs1 fc0 sc0 ls0 ws0">态数据的高层次<span class="_ _2"></span>特征相似性，以<span class="_ _2"></span>及目标模态上<span class="_ _2"></span>的局部相似性。<span class="_ _2"></span>本文通过大量<span class="_ _2"></span>的实验证明了所<span class="_ _2"></span>提出</div></div><div class="pi" data-data='{"ctm":[1.611792,0.000000,0.000000,1.611792,0.000000,0.000000]}'></div></div>
<div id="pf4" class="pf w0 h0" data-page-no="4"><div class="pc pc4 w0 h0"><img class="bi x0 y0 w1 h1" alt="" src="https://csdnimg.cn/release/download_crawler_static/10882440/bg4.jpg"><div class="t m0 x1 h2 y1 ff1 fs0 fc0 sc0 ls0 ws0">CVPR-<span class="ls1">2018</span> <span class="_ _0"> </span><span class="ff2">阿里巴巴收录论文</span> </div><div class="t m0 x2 h3 y2 ff1 fs0 fc0 sc0 ls0 ws0"> </div><div class="t m0 x2 h4 y3 ff2 fs1 fc0 sc0 ls0 ws0">的方法可以准确<span class="_ _2"></span>地匹配图像和文<span class="_ _2"></span>本，并且在<span class="_ _4"> </span><span class="ff3">MSCOCO<span class="_ _4"> </span></span>以及<span class="_ _0"> </span><span class="ff3">Flickr30K<span class="_ _4"> </span></span>的数据集上都取得<span class="_ _2"></span>了<span class="_ _0"> </span><span class="ff3">state-</span></div><div class="t m0 x2 h4 y50 ff3 fs1 fc0 sc0 ls0 ws0">of-the-art<span class="_ _4"> </span><span class="ff2">的效果。<span class="_ _2"></span><span class="ff3"> </span></span></div><div class="t m0 x3 h4 y32 ff2 fs1 fc0 sc0 ls0 ws0">在论文《整体还<span class="_ _2"></span>是局部？应用<span class="_ _4"> </span><span class="ff3">Localized GAN<span class="_ _4"> </span></span>进行图像内容编<span class="_ _2"></span>辑、半监督训<span class="_ _2"></span>练和解决<span class="_ _0"> </span><span class="ff3">mode </span></div><div class="t m0 x2 h4 y33 ff3 fs1 fc0 sc0 ls0 ws0">collapse<span class="_ _4"> </span><span class="ff2">问题》中，作者<span class="_ _2"></span>建立了<span class="_ _4"> </span><span class="ff3">GAN<span class="_ _0"> </span></span>和半监督机器学习<span class="_ _2"></span>中<span class="_ _0"> </span><span class="ff3">Laplace-Bel<span class="_ _2"></span>trami<span class="_ _4"> </span><span class="ff2">算子的联系，在用少</span></span></span></div><div class="t m0 x2 h4 y34 ff2 fs1 fc0 sc0 ls0 ws0">量标注样本训练<span class="_ _2"></span>深度学习模型上<span class="_ _2"></span>取得了优异的<span class="_ _2"></span>性能。同时论文<span class="_ _2"></span>还展示了用<span class="_ _4"> </span><span class="ff3">Localized GA<span class="_ _2"></span>N (LGAN)</span></div><div class="t m0 x2 h4 y35 ff2 fs1 fc0 sc0 ls0 ws0">对给定图像在局<span class="_ _2"></span>部坐标系下进行<span class="_ _2"></span>编辑修改，从<span class="_ _2"></span>而获得具有不同<span class="_ _2"></span>角度、姿态和<span class="_ _2"></span>风格的新图像；<span class="_ _2"></span>以及</div><div class="t m0 x2 h4 y36 ff2 fs1 fc0 sc0 ls0 ws0">如何从流型切向<span class="_ _2"></span>量独立性的角度<span class="_ _2"></span>来解释和解<span class="_ _2"></span>决<span class="_ _0"> </span><span class="ff3">GAN<span class="_ _4"> </span></span>的<span class="_ _0"> </span><span class="ff3">mode collapse<span class="_ _4"> </span></span>问题。<span class="ff3"> </span></div><div class="t m0 x3 h4 y37 ff2 fs1 fc0 sc0 ls0 ws0">论文《处理多种<span class="_ _2"></span>退化类型的卷积<span class="_ _2"></span>超分辨率》针<span class="_ _2"></span>对现有基于<span class="_ _4"> </span><span class="ff3">CNN<span class="_ _0"> </span></span>的单图超分<span class="ff3">(S<span class="_ _2"></span>ISR)<span class="ff2">算法不<span class="_ _2"></span>能扩</span></span></div><div class="t m0 x2 h4 y38 ff2 fs1 fc0 sc0 ls0 ws0">展到用单一模型<span class="_ _2"></span>解决多种不同的<span class="_ _2"></span>图像退化类型<span class="_ _2"></span>的问题，提出了<span class="_ _2"></span>一种维度拉伸<span class="_ _2"></span>策略，使得单个<span class="_ _2"></span>卷积</div><div class="t m0 x2 h4 y39 ff2 fs1 fc0 sc0 ls0 ws0">超分辨率网络能<span class="_ _2"></span>够将<span class="_ _4"> </span><span class="ff3">SISR<span class="_ _0"> </span></span>退化过程的两个<span class="_ _2"></span>关键因素（即模<span class="_ _2"></span>糊核和噪声水<span class="_ _2"></span>平）作为网络输<span class="_ _2"></span>入来解决</div><div class="t m0 x2 h4 y3a ff2 fs1 fc0 sc0 ls0 ws0">这个问题。实验<span class="_ _2"></span>结果表明提出的<span class="_ _2"></span>卷积超分辨率<span class="_ _2"></span>网络可以快速、<span class="_ _2"></span>有效的处理多<span class="_ _2"></span>种图像退化类型<span class="_ _2"></span>，为</div><div class="t m0 x2 h4 y51 ff3 fs1 fc0 sc0 ls0 ws0">SISR<span class="_ _4"> </span><span class="ff2">实际应用提供了一种<span class="_ _2"></span>高效、可扩展的<span class="_ _2"></span>解决方案。<span class="ff3"> </span></span></div><div class="t m0 x3 h4 y52 ff2 fs1 fc0 sc0 ls0 ws0">论文《于尺度空<span class="_ _2"></span>间变换的本征图<span class="_ _2"></span>像分解》将把<span class="_ _2"></span>图像分解为其本<span class="_ _2"></span>征的反射图像<span class="_ _2"></span>和光照图像看作</div><div class="t m0 x2 h4 y53 ff2 fs1 fc0 sc0 ls0 ws0">是一个图像到图<span class="_ _2"></span>像的转换问题，<span class="_ _2"></span>并且将输入和<span class="_ _2"></span>输出在尺度空间<span class="_ _2"></span>进行分解。通<span class="_ _2"></span>过将输出图像（<span class="_ _2"></span>反射</div><div class="t m0 x2 h4 y54 ff2 fs1 fc0 sc0 ls0 ws0">图像和光照图像<span class="_ _2"></span>）扩展到它们的<span class="_ _2"></span>拉普拉斯金字<span class="_ _2"></span>塔的各个成分，<span class="_ _2"></span>论文提出了一<span class="_ _2"></span>种多通道网络结<span class="_ _2"></span>构，</div><div class="t m0 x2 h4 y55 ff2 fs1 fc0 sc0 ls0 ws0">可以在每个通道<span class="_ _2"></span>内并行地学习到<span class="_ _2"></span>一个图像到图<span class="_ _2"></span>像转换函数，这<span class="_ _2"></span>个函数通过一<span class="_ _2"></span>个具有跳过连接<span class="_ _2"></span>的卷</div><div class="t m0 x2 h4 y40 ff2 fs1 fc0 sc0 ls0 ws0">积神经网络来表<span class="_ _2"></span>示。在<span class="_ _4"> </span><span class="ff3 ls2">MPI<span class="_ _1"></span><span class="ls0">-Sintel<span class="_ _4"> </span></span></span>数据集和<span class="_ _0"> </span><span class="ff3">MIT In<span class="_ _2"></span>trinsic Image<span class="_ _2"></span>s<span class="_ _0"> </span><span class="ff2">数据集上<span class="_ _2"></span>结果表明，新提<span class="_ _2"></span>出</span></span></div><div class="t m0 x2 h4 y56 ff2 fs1 fc0 sc0 ls0 ws0">的模型在比之前<span class="_ _2"></span>最先进的技术上<span class="_ _2"></span>有了明显的进<span class="_ _2"></span>步。<span class="ff3"> </span></div><div class="t m0 x3 h4 y57 ff2 fs1 fc0 sc0 ls0 ws0">大多数现有的零<span class="_ _2"></span>样本学习（<span class="ff3">Zero<span class="_ _2"></span>-Shot Learni<span class="_ _2"></span>ng<span class="ff2">，</span>ZSL<span class="ff2">）方<span class="_ _2"></span>法都存在强偏<span class="_ _2"></span>问题。在论文《<span class="_ _2"></span>基于</span></span></div><div class="t m0 x2 h4 y58 ff2 fs1 fc0 sc0 ls0 ws0">直推式无偏嵌入<span class="_ _2"></span>的零样本学习》<span class="_ _2"></span>中，作者提出<span class="_ _2"></span>了一个简单而有<span class="_ _2"></span>效的方法，称<span class="_ _2"></span>为准完全监督学<span class="_ _2"></span>习</div><div class="t m0 x2 h4 y44 ff2 fs1 fc0 sc0 ls0 ws0">（<span class="ff3">QFSL</span>），来缓<span class="_ _2"></span>解此问题。假定<span class="_ _2"></span>标记的源图像<span class="_ _2"></span>和未标记的目标<span class="_ _2"></span>图像都可用于<span class="_ _2"></span>训练。在语义嵌<span class="_ _2"></span>入空</div><div class="t m0 x2 h4 y45 ff2 fs1 fc0 sc0 ls0 ws0">间中，被标记的<span class="_ _2"></span>源图像被映射到<span class="_ _2"></span>由源类别指定<span class="_ _2"></span>的若干个嵌入点<span class="_ _2"></span>，并且未标记<span class="_ _2"></span>的目标图像被强<span class="_ _2"></span>制映</div><div class="t m0 x2 h4 y46 ff2 fs1 fc0 sc0 ls0 ws0">射到由目标类别<span class="_ _2"></span>指定的其他点。<span class="_ _2"></span>在<span class="_ _0"> </span><span class="ff3">AwA2</span>，<span class="ff3">CUB<span class="_ _4"> </span></span>和<span class="_ _0"> </span><span class="ff3 ls2">SUN<span class="_ _0"> </span></span>数据集上进行的实<span class="_ _2"></span>验表明，文章的<span class="_ _2"></span>方法在遵</div><div class="t m0 x2 h4 y47 ff2 fs1 fc0 sc0 ls0 ws0">循广义<span class="_ _4"> </span><span class="ff3">ZSL<span class="_ _0"> </span></span>设置的情况下<span class="_ _2"></span>比现有技术的方法优<span class="_ _2"></span>越。<span class="ff3"> </span></div><div class="t m0 x3 h4 y48 ff2 fs1 fc0 sc0 ls0 ws0">当下计算机视觉<span class="_ _2"></span>技术无疑是<span class="_ _4"> </span><span class="ff3">AI<span class="_ _0"> </span></span>浪潮中火热的<span class="_ _2"></span>题目，受关注<span class="_ _2"></span>的程度持续升<span class="_ _2"></span>温。视觉技术的<span class="_ _2"></span>渗</div><div class="t m0 x2 h4 y49 ff2 fs1 fc0 sc0 ls0 ws0">透，既可能是对<span class="_ _2"></span>传统商业的改造<span class="_ _2"></span>使之看到新的<span class="_ _2"></span>商业机会，还可<span class="_ _2"></span>能是创造了全<span class="_ _2"></span>新的商业需求和<span class="_ _2"></span>市</div><div class="t m0 x2 h4 y4a ff2 fs1 fc0 sc0 ls0 ws0">场。好的视觉技<span class="_ _2"></span>术不仅需要有好<span class="_ _2"></span>的方法指引，<span class="_ _2"></span>而且需要在实际<span class="_ _2"></span>的场景中形成<span class="_ _2"></span>数据闭环和不断<span class="_ _2"></span>打</div><div class="t m0 x2 h4 y4b ff2 fs1 fc0 sc0 ls0 ws0">磨。未来的计算<span class="_ _2"></span>机视觉技术一定<span class="_ _2"></span>是理论探索和<span class="_ _2"></span>数据实践的共同<span class="_ _2"></span>推进。希望这<span class="_ _2"></span>本论文合集能抛<span class="_ _2"></span>砖引</div><div class="t m0 x2 h4 y4c ff2 fs1 fc0 sc0 ls0 ws0">玉，给学术界和<span class="_ _2"></span>工业界带来一些<span class="_ _2"></span>输入，共同推<span class="_ _2"></span>进计算机视觉技<span class="_ _2"></span>术的发展。<span class="_ _2"></span><span class="ff3"> </span></div><div class="t m0 xc h4 y59 ff2 fs1 fc0 sc0 ls0 ws0">阿里巴巴资深算<span class="_ _2"></span>法专家<span class="ff3"> </span><span class="ls2">潘攀</span>（启磐）<span class="_ _2"></span><span class="ff3"> </span></div><div class="t m0 xd h4 y5a ff3 fs1 fc0 sc0 ls0 ws0">2018<span class="_ _4"> </span><span class="ff2">年<span class="_ _0"> </span></span>12<span class="_ _0"> </span><span class="ff2">月</span> <span class="ff2">于北京</span> </div><div class="t m0 x2 h4 y5b ff3 fs1 fc0 sc0 ls0 ws0"> </div></div><div class="pi" data-data='{"ctm":[1.611792,0.000000,0.000000,1.611792,0.000000,0.000000]}'></div></div>
<div id="pf5" class="pf w0 h0" data-page-no="5"><div class="pc pc5 w0 h0"><img class="bi x0 y0 w1 h1" alt="" src="https://csdnimg.cn/release/download_crawler_static/10882440/bg5.jpg"><div class="t m0 x1 h2 y1 ff1 fs0 fc0 sc0 ls0 ws0">CVPR-<span class="ls1">2018</span> <span class="_ _0"> </span><span class="ff2">阿里巴巴收录论文</span> </div><div class="t m0 x2 h3 y2 ff1 fs0 fc0 sc0 ls0 ws0"> </div><div class="t m0 xe hc y5c ff2 fs8 fc4 sc2 ls3 ws0">目录<span class="ff5 ls0"> </span></div><div class="t m0 x2 hd y5d ff1 fs6 fc0 sc0 ls0 ws0"> </div><div class="t m0 x2 he y5e ff1 fs6 fc0 sc0 ls0 ws0">CVPR-<span class="ls4">18<span class="_ _0"> </span></span><span class="ff2">阿里巴巴<span class="_ _0"> </span></span>Spotlight<span class="_"> </span><span class="ff2">论文：基于时间尺度选择的在线<span class="_ _2"></span>行为预测<span class="ff1"> <span class="ls5">.........................................</span> <span class="_ _5"></span>1<span class="fs1"> </span></span></span></div><div class="t m0 x2 he y5f ff1 fs6 fc0 sc0 ls0 ws0">CVPR-<span class="ls4">18<span class="_ _0"> </span></span><span class="ff2">阿里巴巴<span class="_ _0"> </span></span>Spotlight<span class="_"> </span><span class="ff2">论文：基于语境对比特征和门控<span class="_ _2"></span>多尺度融合的场<span class="_ _2"></span>景分割<span class="ff1"> <span class="ls5">.................</span> <span class="_ _6"></span>4<span class="fs1"> </span></span></span></div><div class="t m0 xf he y60 ff1 fs6 fc0 sc0 ls6 ws0">1.<span class="fs1 ls0"> <span class="_ _7"> </span><span class="ff2 fs6">语境对比局部特征<span class="_ _2"></span><span class="ff1"> <span class="_ _8"></span><span class="ls5">.....................................................................................................................<span class="ls0"> <span class="_ _5"></span>5<span class="fs1"> </span></span></span></span></span></span></div><div class="t m0 xf he y1b ff1 fs6 fc0 sc0 ls6 ws0">2.<span class="fs1 ls0"> <span class="_ _7"> </span><span class="ff2 fs6">门控多尺度融合<span class="ff1"> <span class="_ _6"></span><span class="ls5">.........................................................................................................................<span class="ls0"> <span class="_ _5"></span>6<span class="fs1"> </span></span></span></span></span></span></div><div class="t m0 xf he y61 ff1 fs6 fc0 sc0 ls6 ws0">3.<span class="fs1 ls0"> <span class="_ _7"> </span><span class="ff2 fs6">实验<span class="ff1"> <span class="_ _6"></span><span class="ls5">.............................................................................................................................................<span class="ls0"> <span class="_ _5"></span>7<span class="fs1"> </span></span></span></span></span></span></div><div class="t m0 x2 he y62 ff1 fs6 fc0 sc0 ls0 ws0">CVPR-<span class="ls4">18<span class="_ _0"> </span></span><span class="ff2">阿里巴巴<span class="_ _0"> </span></span>Spotlight<span class="_"> </span><span class="ff2">论文：所见所想所找－基于生成<span class="_ _2"></span>模型的跨模态检<span class="_ _2"></span>索<span class="ff1"> <span class="ls5">.........................</span> <span class="_ _6"></span>8<span class="fs1"> </span></span></span></div><div class="t m0 xf he y63 ff1 fs6 fc0 sc0 ls6 ws0">1.<span class="fs1 ls0"> <span class="_ _7"> </span><span class="ff2 fs6">摘要<span class="ff1"> <span class="_ _6"></span><span class="ls5">.............................................................................................................................................<span class="ls0"> <span class="_ _5"></span>8<span class="fs1"> </span></span></span></span></span></span></div><div class="t m0 xf he y64 ff1 fs6 fc0 sc0 ls6 ws0">2.<span class="fs1 ls0"> <span class="_ _7"> </span><span class="ff2 fs6">引言<span class="ff1"> <span class="_ _6"></span><span class="ls5">.............................................................................................................................................<span class="ls0"> <span class="_ _5"></span>8<span class="fs1"> </span></span></span></span></span></span></div><div class="t m0 xf he y65 ff1 fs6 fc0 sc0 ls6 ws0">3.<span class="fs1 ls0"> <span class="_ _7"> </span><span class="ff2 fs6">方法<span class="ff1"> <span class="_ _6"></span><span class="ls5">.............................................................................................................................................<span class="ls0"> <span class="_ _5"></span>9<span class="fs1"> </span></span></span></span></span></span></div><div class="t m0 xf he y66 ff1 fs6 fc0 sc0 ls6 ws0">4.<span class="fs1 ls0"> <span class="_ _7"> </span><span class="ff2 fs6">实验<span class="ff1"> <span class="_ _6"></span><span class="ls5">...........................................................................................................................................<span class="ls0"> <span class="_ _5"></span><span class="ls6">10<span class="fs1 ls0"> </span></span></span></span></span></span></span></div><div class="t m0 xf he y67 ff1 fs6 fc0 sc0 ls6 ws0">5.<span class="fs1 ls0"> <span class="_ _7"> </span><span class="ff2 fs6">总结<span class="ff1"> <span class="_ _6"></span><span class="ls5">...........................................................................................................................................<span class="ls0"> <span class="_ _5"></span><span class="ls6">11<span class="fs1 ls0"> </span></span></span></span></span></span></span></div><div class="t m0 x2 he y68 ff1 fs6 fc0 sc0 ls0 ws0">CVPR2018<span class="_"> </span><span class="ff2">阿里巴巴<span class="_ _0"> </span></span>Poster<span class="_"> </span><span class="ff2">论文：整体还是局部？应用<span class="_ _0"> </span></span>Localized <span class="_ _2"></span>GAN<span class="_ _0"> </span><span class="ff2">进行图像内容编辑、半监</span></div><div class="t m0 x2 he y69 ff2 fs6 fc0 sc0 ls0 ws0">督训练和解决<span class="_ _4"> </span><span class="ff1">m<span class="_ _1"></span>ode colla<span class="_ _2"></span>pse<span class="_ _0"> </span><span class="ff2">问题</span> <span class="_ _8"></span><span class="ls5">.....................................................................................................<span class="ls0"> <span class="_ _5"></span><span class="ls6">11<span class="fs1 ls0"> </span></span></span></span></span></div><div class="t m0 xf he y6a ff1 fs6 fc0 sc0 ls6 ws0">1.<span class="fs1 ls0"> <span class="_ _7"> </span><span class="ff2 fs6">摘要<span class="ff1"> <span class="_ _6"></span><span class="ls5">...........................................................................................................................................<span class="ls0"> <span class="_ _5"></span><span class="ls6">11<span class="fs1 ls0"> </span></span></span></span></span></span></span></div><div class="t m0 xf he y6b ff1 fs6 fc0 sc0 ls6 ws0">2.<span class="fs1 ls0"> <span class="_ _7"> </span><span class="fs6">GAN<span class="_"> </span><span class="ff2">和基于图模型的半监督机器学<span class="_ _2"></span>习的关系<span class="ff1"> <span class="_ _5"></span><span class="ls5">.......................................................................<span class="ls0"> <span class="_ _5"></span><span class="ls6">12<span class="fs1 ls0"> </span></span></span></span></span></span></span></span></div><div class="t m0 xf he y6c ff1 fs6 fc0 sc0 ls6 ws0">3.<span class="fs1 ls0"> <span class="_ _7"> </span><span class="ff2 fs6">用全局还是局部坐标来<span class="_ _2"></span>研究<span class="_ _0"> </span><span class="ff1">GAN <span class="_ _9"></span><span class="ls5">...........................................................................................<span class="ls0"> <span class="_ _5"></span><span class="ls6">14<span class="fs1 ls0"> </span></span></span></span></span></span></span></div><div class="t m0 xf he y6d ff1 fs6 fc0 sc0 ls6 ws0">4.<span class="fs1 ls0"> <span class="_ _7"> </span><span class="ff2 fs6">从几何角度研究<span class="_ _4"> </span><span class="ff1">Mode collapse<span class="_ _0"> </span></span>问题<span class="ff1"> <span class="ls5">.....................................................................................</span> <span class="_ _6"></span><span class="ls6">15<span class="fs1 ls0"> </span></span></span></span></span></div><div class="t m0 x2 he y6e ff1 fs6 fc0 sc0 ls0 ws0">CVPR2018<span class="_"> </span><span class="ff2">阿里巴巴<span class="_ _0"> </span></span>Poster<span class="_"> </span><span class="ff2">论文：处理多种退化类型的卷积<span class="_ _2"></span>超分辨率<span class="ff1"> <span class="_ _a"></span><span class="ls5">.........................................<span class="ls0"> <span class="_ _5"></span><span class="ls6">16<span class="fs1 ls0"> </span></span></span></span></span></span></div><div class="t m0 xf he y6f ff1 fs6 fc0 sc0 ls0 ws0">1<span class="ff2">．摘要</span> <span class="ls5">............................................................................................................................................</span> <span class="_ _5"></span><span class="ls6">16<span class="fs1 ls0"> </span></span></div><div class="t m0 xf he y70 ff1 fs6 fc0 sc0 ls0 ws0">2<span class="ff2">．引言</span> <span class="ls5">............................................................................................................................................</span> <span class="_ _5"></span><span class="ls6">16<span class="fs1 ls0"> </span></span></div><div class="t m0 xf he y71 ff1 fs6 fc0 sc0 ls0 ws0">3<span class="ff2">．方法</span> <span class="ls5">............................................................................................................................................</span> <span class="_ _5"></span><span class="ls6">17<span class="fs1 ls0"> </span></span></div><div class="t m0 xf he y72 ff1 fs6 fc0 sc0 ls0 ws0">4<span class="ff2">．实验</span> <span class="ls5">............................................................................................................................................</span> <span class="_ _5"></span><span class="ls6">19<span class="fs1 ls0"> </span></span></div><div class="t m0 xf he y73 ff1 fs6 fc0 sc0 ls0 ws0">5<span class="ff2">．结论</span> <span class="ls5">............................................................................................................................................</span> <span class="_ _5"></span><span class="ls6">20<span class="fs1 ls0"> </span></span></div><div class="t m0 x2 he y74 ff1 fs6 fc0 sc0 ls0 ws0">CVPR2018<span class="_"> </span><span class="ff2">阿里巴巴<span class="_ _0"> </span></span>Poster<span class="_"> </span><span class="ff2">论文：基于尺度空间变换的本征<span class="_ _2"></span>图像分解<span class="ff1"> <span class="_ _a"></span><span class="ls5">.........................................<span class="ls0"> <span class="_ _5"></span><span class="ls6">21<span class="fs1 ls0"> </span></span></span></span></span></span></div><div class="t m0 xf he y75 ff2 fs6 fc0 sc0 ls0 ws0">摘要<span class="ff1"> <span class="ls5">..................................................................................................................................................</span> <span class="_ _5"></span><span class="ls6">21<span class="fs1 ls0"> </span></span></span></div><div class="t m0 xf he y76 ff3 fs6 fc0 sc0 ls0 ws0">1. <span class="ff2">引言<span class="ff1"> <span class="ls5">............................................................................................................................................</span> <span class="_ _5"></span><span class="ls6">21<span class="fs1 ls0"> </span></span></span></span></div><div class="t m0 xf he y77 ff3 fs6 fc0 sc0 ls0 ws0">2. <span class="ff2">相关工作（略）<span class="_ _2"></span><span class="ff1"> <span class="_ _a"></span><span class="ls5">.........................................................................................................................<span class="ls0"> <span class="_ _5"></span><span class="ls6">22<span class="fs1 ls0"> </span></span></span></span></span></span></div><div class="t m0 xf he y78 ff3 fs6 fc0 sc0 ls0 ws0">3. <span class="ff2">我们的方法<span class="ff1 ls5">................................................................................................................................<span class="ls0">. <span class="_ _9"></span><span class="ls6">22<span class="fs1 ls0"> </span></span></span></span></span></div><div class="t m0 x10 he y79 ff3 fs6 fc0 sc0 ls0 ws0">3.1 <span class="ff2">网络结构的演化<span class="_ _2"></span><span class="ff1"> <span class="_ _a"></span><span class="ls5">...................................................................................................................<span class="ls0"> <span class="_ _5"></span><span class="ls6">23<span class="fs1 ls0"> </span></span></span></span></span></span></div><div class="t m0 x10 he y7a ff3 fs6 fc0 sc0 ls0 ws0">3.2 <span class="ff2">残差块<span class="ff1"> <span class="_ _b"></span><span class="ls5">...................................................................................................................................<span class="ls0"> <span class="_ _5"></span><span class="ls6">24<span class="fs1 ls0"> </span></span></span></span></span></span></div><div class="t m0 x10 he y7b ff3 fs6 fc0 sc0 ls0 ws0">3.3 <span class="ff2">损失函数<span class="ff1"> <span class="_ _b"></span><span class="ls5">...............................................................................................................................<span class="ls0"> <span class="_ _5"></span><span class="ls6">24<span class="fs1 ls0"> </span></span></span></span></span></span></div><div class="t m0 x10 he y7c ff3 fs6 fc0 sc0 ls0 ws0">3.4 <span class="ff2">数据增强训练<span class="_ _2"></span><span class="ff1"> <span class="_ _a"></span><span class="ls5">.......................................................................................................................<span class="ls0"> <span class="_ _5"></span><span class="ls6">25<span class="fs1 ls0"> </span></span></span></span></span></span></div><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m1"></div></a></div><div class="pi" data-data='{"ctm":[1.611792,0.000000,0.000000,1.611792,0.000000,0.000000]}'></div></div>
