计算机视觉目标检测yolov4文件 yolov4.7z

  • O4_991753
    了解作者
  • 382.7MB
    文件大小
  • 7z
    文件格式
  • 0
    收藏次数
  • VIP专享
    资源类型
  • 0
    下载次数
  • 2022-05-07 03:43
    上传日期
计算机视觉目标检测yolov4文件。
yolov4.7z
  • YOLOv4.pdf
    3.8MB
  • yolov4.conv.137
    162.2MB
  • yolov4.cfg
    11.9KB
  • darknet-master.zip
    7.8MB
  • yolov4.weights
    245.8MB
内容介绍
<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://static.pudn.com/base/css/base.min.css"> <link rel="stylesheet" href="https://static.pudn.com/base/css/fancy.min.css"> <link rel="stylesheet" href="https://static.pudn.com/prod/directory_preview_static/6275eafd16f2c0769caad825/raw.css"> <script src="https://static.pudn.com/base/js/compatibility.min.js"></script> <script src="https://static.pudn.com/base/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://static.pudn.com/prod/directory_preview_static/6275eafd16f2c0769caad825/bg1.jpg"><div class="t m0 x1 h2 y1 ff1 fs0 fc0 sc0 ls0 ws0">Y<span class="_ _0"></span>OLOv4:<span class="_ _1"> </span>Optimal<span class="_"> </span>Speed<span class="_"> </span>and<span class="_"> </span>Accuracy<span class="_"> </span>of<span class="_"> </span>Object<span class="_"> </span>Detection</div><div class="t m0 x2 h3 y2 ff2 fs1 fc0 sc0 ls0 ws0">Alex<span class="_ _0"></span>ey<span class="_"> </span>Bochko<span class="_ _0"></span>vskiy</div><div class="t m0 x3 h4 y3 ff3 fs2 fc0 sc0 ls0 ws0">&#8727;</div><div class="t m0 x4 h5 y4 ff4 fs3 fc0 sc0 ls0 ws0">alexeyab84@gmail.com</div><div class="t m0 x5 h3 y2 ff2 fs1 fc0 sc0 ls0 ws0">Chien-Y<span class="_ _2"></span>ao<span class="_"> </span>W<span class="_ _2"></span>ang</div><div class="t m0 x6 h4 y3 ff3 fs2 fc0 sc0 ls0 ws0">&#8727;</div><div class="t m0 x7 h3 y4 ff2 fs1 fc0 sc0 ls0 ws0">Institute<span class="_"> </span>of<span class="_"> </span>Information<span class="_"> </span>Science</div><div class="t m0 x8 h3 y5 ff2 fs1 fc0 sc0 ls0 ws0">Academia<span class="_"> </span>Sinica,<span class="_"> </span>T<span class="_ _2"></span>aiwan</div><div class="t m0 x9 h5 y6 ff4 fs3 fc0 sc0 ls0 ws0">kinyiu@iis.sinica.edu.tw</div><div class="t m0 xa h3 y2 ff2 fs1 fc0 sc0 ls0 ws0">Hong-Y<span class="_ _2"></span>uan<span class="_"> </span>Mark<span class="_"> </span>Liao</div><div class="t m0 xb h3 y7 ff2 fs1 fc0 sc0 ls0 ws0">Institute<span class="_"> </span>of<span class="_"> </span>Information<span class="_"> </span>Science</div><div class="t m0 xc h3 y5 ff2 fs1 fc0 sc0 ls0 ws0">Academia<span class="_"> </span>Sinica,<span class="_"> </span>T<span class="_ _2"></span>aiwan</div><div class="t m0 xd h5 y6 ff4 fs3 fc0 sc0 ls0 ws0">liao@iis.sinica.edu.tw</div><div class="t m0 xe h6 y8 ff1 fs1 fc0 sc0 ls0 ws0">Abstract</div><div class="t m0 xf h7 y9 ff5 fs4 fc0 sc0 ls0 ws0">Ther<span class="_ _0"></span>e<span class="_ _3"> </span>ar<span class="_ _0"></span>e<span class="_ _3"> </span>a<span class="_ _3"> </span>huge<span class="_ _4"> </span>number<span class="_ _4"> </span>of<span class="_ _3"> </span>featur<span class="_ _0"></span>es<span class="_ _3"> </span>which<span class="_ _4"> </span>are<span class="_ _4"> </span>said<span class="_ _4"> </span>to</div><div class="t m0 x10 h7 ya ff5 fs4 fc0 sc0 ls0 ws0">impr<span class="_ _0"></span>ove<span class="_ _1"> </span>Con<span class="_ _0"></span>volutional<span class="_ _5"> </span>Neural<span class="_ _5"> </span>Network<span class="_ _5"> </span>(CNN)<span class="_ _5"> </span>accuracy<span class="_ _0"></span>.</div><div class="t m0 x10 h7 yb ff5 fs4 fc0 sc0 ls0 ws0">Practical<span class="_ _6"> </span>testing<span class="_ _6"> </span>of<span class="_ _4"> </span>combinations<span class="_ _6"> </span>of<span class="_ _6"> </span>such<span class="_ _6"> </span>features<span class="_ _6"> </span>on<span class="_ _6"> </span>larg<span class="_ _0"></span>e</div><div class="t m0 x10 h7 yc ff5 fs4 fc0 sc0 ls0 ws0">datasets,<span class="_ _1"> </span>and<span class="_ _5"> </span>theor<span class="_ _0"></span>etical<span class="_ _5"> </span>justi&#64257;cation<span class="_ _5"> </span>of<span class="_ _5"> </span>the<span class="_ _5"> </span>r<span class="_ _0"></span>esult,<span class="_ _1"> </span>is<span class="_ _5"> </span>r<span class="_ _0"></span>e-</div><div class="t m0 x10 h7 yd ff5 fs4 fc0 sc0 ls0 ws0">quir<span class="_ _0"></span>ed.<span class="_ _4"> </span>Some<span class="_ _7"> </span>featur<span class="_ _0"></span>es<span class="_ _7"> </span>operate<span class="_ _7"> </span>on<span class="_ _7"> </span>certain<span class="_ _8"> </span>models<span class="_ _7"> </span>exclusively</div><div class="t m0 x10 h7 ye ff5 fs4 fc0 sc0 ls0 ws0">and<span class="_ _7"> </span>for<span class="_ _8"> </span>certain<span class="_ _7"> </span>pr<span class="_ _0"></span>oblems<span class="_"> </span>e<span class="_ _0"></span>xclusively<span class="_ _0"></span>,<span class="_ _8"> </span>or<span class="_ _7"> </span>only<span class="_ _8"> </span>for<span class="_ _7"> </span>small-scale</div><div class="t m0 x10 h7 yf ff5 fs4 fc0 sc0 ls0 ws0">datasets;<span class="_"> </span>while<span class="_ _6"> </span>some<span class="_"> </span>features,<span class="_"> </span>such<span class="_"> </span>as<span class="_"> </span>batch-normalization</div><div class="t m0 x10 h7 y10 ff5 fs4 fc0 sc0 ls0 ws0">and<span class="_"> </span>residual-connections,<span class="_"> </span>are<span class="_"> </span>applicable<span class="_ _6"> </span>to<span class="_ _6"> </span>the<span class="_ _6"> </span>majority<span class="_ _6"> </span>of</div><div class="t m0 x10 h7 y11 ff5 fs4 fc0 sc0 ls0 ws0">models,<span class="_"> </span>tasks,<span class="_"> </span>and<span class="_ _6"> </span>datasets.<span class="_ _4"> </span>W<span class="_ _0"></span>e<span class="_"> </span>assume<span class="_"> </span>that<span class="_"> </span>such<span class="_"> </span>universal</div><div class="t m0 x10 h7 y12 ff5 fs4 fc0 sc0 ls0 ws0">featur<span class="_ _0"></span>es<span class="_ _9"> </span>include<span class="_ _9"> </span>W<span class="_ _2"></span>eighted-Residual-Connections<span class="_ _9"> </span>(WRC),</div><div class="t m0 x10 h7 y13 ff5 fs4 fc0 sc0 ls0 ws0">Cr<span class="_ _0"></span>oss-Stage-P<span class="_ _a"></span>artial-connections<span class="_ _5"> </span>(CSP),<span class="_ _5"> </span>Cr<span class="_ _0"></span>oss<span class="_ _1"> </span>mini-Batch</div><div class="t m0 x10 h7 y14 ff5 fs4 fc0 sc0 ls0 ws0">Normalization<span class="_ _b"> </span>(CmBN),<span class="_ _b"> </span>Self-adversarial-tr<span class="_ _0"></span>aining<span class="_ _b"> </span>(SAT)</div><div class="t m0 x10 h7 y15 ff5 fs4 fc0 sc0 ls0 ws0">and<span class="_ _5"> </span>Mish-activation.<span class="_ _c"> </span>W<span class="_ _a"></span>e<span class="_ _5"> </span>use<span class="_ _1"> </span>new<span class="_ _5"> </span>features:<span class="_ _d"> </span>WRC,<span class="_ _5"> </span>CSP<span class="_ _2"></span>,</div><div class="t m0 x10 h7 y16 ff5 fs4 fc0 sc0 ls0 ws0">CmBN,<span class="_ _3"> </span>SAT<span class="_ _a"></span>,<span class="_ _e"> </span>Mish<span class="_ _3"> </span>activation,<span class="_ _5"> </span>Mosaic<span class="_ _e"> </span>data<span class="_ _e"> </span>augmentation,</div><div class="t m0 x10 h7 y17 ff5 fs4 fc0 sc0 ls0 ws0">CmBN,<span class="_ _8"> </span>Dr<span class="_ _0"></span>opBlock<span class="_ _7"> </span>re<span class="_ _0"></span>gularization,<span class="_ _8"> </span>and<span class="_ _8"> </span>CIoU<span class="_"> </span>loss,<span class="_ _7"> </span>and<span class="_"> </span>com-</div><div class="t m0 x10 h7 y18 ff5 fs4 fc0 sc0 ls0 ws0">bine<span class="_ _8"> </span>some<span class="_ _8"> </span>of<span class="_ _8"> </span>them<span class="_ _8"> </span>to<span class="_ _8"> </span>achie<span class="_ _0"></span>ve<span class="_"> </span>state-of-the-art<span class="_ _7"> </span>results:<span class="_"> </span>43.5%</div><div class="t m0 x10 h7 y19 ff5 fs4 fc0 sc0 ls0 ws0">AP<span class="_ _1"> </span>(65.7%<span class="_ _f"> </span>AP</div><div class="t m0 x11 h8 y1a ff6 fs5 fc0 sc0 ls0 ws0">50</div><div class="t m0 x12 h7 y19 ff5 fs4 fc0 sc0 ls0 ws0">)<span class="_ _1"> </span>for<span class="_ _f"> </span>the<span class="_ _1"> </span>MS<span class="_ _f"> </span>COCO<span class="_ _1"> </span>dataset<span class="_ _f"> </span>at<span class="_ _1"> </span>a<span class="_ _f"> </span>r<span class="_ _0"></span>eal-</div><div class="t m0 x10 h9 y1b ff5 fs4 fc0 sc0 ls0 ws0">time<span class="_ _4"> </span>speed<span class="_ _3"> </span>of<span class="_ _3"> </span><span class="ff7">&#8764;</span>65<span class="_ _4"> </span>FPS<span class="_ _3"> </span>on<span class="_ _3"> </span>T<span class="_ _2"></span>esla<span class="_ _3"> </span>V100.<span class="_ _9"> </span>Sour<span class="_ _0"></span>ce<span class="_ _3"> </span>code<span class="_ _4"> </span>is<span class="_ _3"> </span>at</div><div class="t m0 x10 h7 y1c ff8 fs4 fc1 sc0 ls0 ws0">https://github.com/AlexeyAB/darknet<span class="ff5 fc0">.</span></div><div class="t m0 x10 h6 y1d ff1 fs1 fc0 sc0 ls0 ws0">1.<span class="_"> </span>Introduction</div><div class="t m0 xf ha y1e ff2 fs4 fc0 sc0 ls0 ws0">The<span class="_ _6"> </span>majority<span class="_ _6"> </span>of<span class="_ _6"> </span>CNN-based<span class="_ _4"> </span>object<span class="_"> </span>detectors<span class="_ _4"> </span>are<span class="_ _6"> </span>largely</div><div class="t m0 x10 ha y1f ff2 fs4 fc0 sc0 ls0 ws0">applicable<span class="_"> </span>only<span class="_"> </span>for<span class="_"> </span>recommendation<span class="_ _8"> </span>systems.<span class="_ _4"> </span>For<span class="_"> </span>e<span class="_ _0"></span>xample,</div><div class="t m0 x10 ha y20 ff2 fs4 fc0 sc0 ls0 ws0">searching<span class="_ _4"> </span>for<span class="_ _4"> </span>free<span class="_ _4"> </span>parking<span class="_ _4"> </span>spaces<span class="_ _4"> </span>via<span class="_ _4"> </span>urban<span class="_ _4"> </span>video<span class="_ _4"> </span>cameras</div><div class="t m0 x10 ha y21 ff2 fs4 fc0 sc0 ls0 ws0">is<span class="_ _6"> </span>ex<span class="_ _0"></span>ecuted<span class="_ _4"> </span>by<span class="_"> </span>slow<span class="_ _6"> </span>accurate<span class="_ _6"> </span>models,<span class="_ _6"> </span>whereas<span class="_ _6"> </span>car<span class="_ _6"> </span>collision</div><div class="t m0 x10 ha y22 ff2 fs4 fc0 sc0 ls0 ws0">warning<span class="_ _5"> </span>is<span class="_ _5"> </span>related<span class="_ _5"> </span>to<span class="_ _5"> </span>fast<span class="_ _5"> </span>inaccurate<span class="_ _1"> </span>models.<span class="_ _10"> </span>Improving</div><div class="t m0 x10 ha y23 ff2 fs4 fc0 sc0 ls0 ws0">the<span class="_ _3"> </span>real-time<span class="_ _3"> </span>object<span class="_ _e"> </span>detector<span class="_ _3"> </span>accuracy<span class="_ _3"> </span>enables<span class="_ _3"> </span>using<span class="_ _e"> </span>them</div><div class="t m0 x10 ha y24 ff2 fs4 fc0 sc0 ls0 ws0">not<span class="_ _4"> </span>only<span class="_ _3"> </span>for<span class="_ _4"> </span>hint<span class="_ _4"> </span>generating<span class="_ _3"> </span>recommendation<span class="_ _4"> </span>systems,<span class="_ _3"> </span>but</div><div class="t m0 x10 ha y25 ff2 fs4 fc0 sc0 ls0 ws0">also<span class="_ _6"> </span>for<span class="_ _6"> </span>stand-alone<span class="_ _4"> </span>process<span class="_"> </span>management<span class="_ _4"> </span>and<span class="_ _6"> </span>human<span class="_ _6"> </span>input</div><div class="t m0 x10 ha y26 ff2 fs4 fc0 sc0 ls0 ws0">reduction.<span class="_ _11"> </span>Real-time<span class="_ _4"> </span>object<span class="_ _4"> </span>detector<span class="_ _4"> </span>operation<span class="_ _3"> </span>on<span class="_ _4"> </span>con<span class="_ _0"></span>ven-</div><div class="t m0 x10 ha y27 ff2 fs4 fc0 sc0 ls0 ws0">tional<span class="_ _4"> </span>Graphics<span class="_ _4"> </span>Processing<span class="_ _4"> </span>Units<span class="_ _3"> </span>(GPU)<span class="_ _4"> </span>allo<span class="_ _0"></span>ws<span class="_ _3"> </span>their<span class="_ _4"> </span>mass</div><div class="t m0 x10 ha y28 ff2 fs4 fc0 sc0 ls0 ws0">usage<span class="_ _e"> </span>at<span class="_ _5"> </span>an<span class="_ _e"> </span>affordable<span class="_ _e"> </span>price.<span class="_ _12"> </span>The<span class="_ _e"> </span>most<span class="_ _5"> </span>accurate<span class="_ _e"> </span>modern</div><div class="t m0 x10 ha y29 ff2 fs4 fc0 sc0 ls0 ws0">neural<span class="_ _7"> </span>networks<span class="_ _7"> </span>do<span class="_ _8"> </span>not<span class="_ _7"> </span>operate<span class="_ _8"> </span>in<span class="_ _7"> </span>real<span class="_ _8"> </span>time<span class="_ _7"> </span>and<span class="_ _8"> </span>require<span class="_ _7"> </span>large</div><div class="t m0 x10 ha y2a ff2 fs4 fc0 sc0 ls0 ws0">number<span class="_ _4"> </span>of<span class="_ _4"> </span>GPUs<span class="_ _4"> </span>for<span class="_ _4"> </span>training<span class="_ _6"> </span>with<span class="_ _4"> </span>a<span class="_ _4"> </span>large<span class="_ _4"> </span>mini-batch-size.</div><div class="t m0 x10 ha y2b ff2 fs4 fc0 sc0 ls0 ws0">W<span class="_ _a"></span>e<span class="_"> </span>address<span class="_"> </span>such<span class="_"> </span>problems<span class="_ _7"> </span>through<span class="_"> </span>creating<span class="_"> </span>a<span class="_"> </span>CNN<span class="_ _8"> </span>that<span class="_"> </span>op-</div><div class="t m0 x10 ha y2c ff2 fs4 fc0 sc0 ls0 ws0">erates<span class="_ _3"> </span>in<span class="_ _3"> </span>real-time<span class="_ _3"> </span>on<span class="_ _3"> </span>a<span class="_ _3"> </span>con<span class="_ _0"></span>ventional<span class="_ _3"> </span>GPU,<span class="_ _3"> </span>and<span class="_ _3"> </span>for<span class="_ _3"> </span>which</div><div class="t m0 x10 ha y2d ff2 fs4 fc0 sc0 ls0 ws0">training<span class="_"> </span>requires<span class="_"> </span>only<span class="_"> </span>one<span class="_"> </span>con<span class="_ _0"></span>ventional<span class="_"> </span>GPU.</div><div class="t m0 x13 ha y2e ff2 fs4 fc0 sc0 ls0 ws0">Figure<span class="_ _6"> </span>1:<span class="_ _5"> </span>Comparison<span class="_ _6"> </span>of<span class="_ _4"> </span>the<span class="_ _6"> </span>proposed<span class="_ _4"> </span>YOLOv4<span class="_"> </span>and<span class="_ _4"> </span>other</div><div class="t m0 x13 ha y2f ff2 fs4 fc0 sc0 ls0 ws0">state-of-the-art<span class="_"> </span>object<span class="_"> </span>detectors.<span class="_ _3"> </span>Y<span class="_ _0"></span>OLOv4<span class="_ _6"> </span>runs<span class="_"> </span>twice<span class="_"> </span>faster</div><div class="t m0 x13 ha y30 ff2 fs4 fc0 sc0 ls0 ws0">than<span class="_ _6"> </span>Ef&#64257;cientDet<span class="_ _6"> </span>with<span class="_ _4"> </span>comparable<span class="_ _6"> </span>performance.<span class="_ _1"> </span>Improves</div><div class="t m0 x13 ha y31 ff2 fs4 fc0 sc0 ls0 ws0">Y<span class="_ _0"></span>OLOv3&#8217;<span class="_ _0"></span>s<span class="_"> </span>AP<span class="_"> </span>and<span class="_"> </span>FPS<span class="_"> </span>by<span class="_"> </span>10%<span class="_"> </span>and<span class="_"> </span>12%,<span class="_"> </span>respectively<span class="_ _a"></span>.</div><div class="t m0 x14 ha y32 ff2 fs4 fc0 sc0 ls0 ws0">The<span class="_"> </span>main<span class="_ _6"> </span>goal<span class="_ _6"> </span>of<span class="_ _6"> </span>this<span class="_"> </span>work<span class="_ _6"> </span>is<span class="_ _6"> </span>designing<span class="_"> </span>a<span class="_ _6"> </span>fast<span class="_ _6"> </span>operating</div><div class="t m0 x13 ha y33 ff2 fs4 fc0 sc0 ls0 ws0">speed<span class="_"> </span>of<span class="_ _6"> </span>an<span class="_"> </span>object<span class="_ _6"> </span>detector<span class="_"> </span>in<span class="_ _6"> </span>production<span class="_ _6"> </span>systems<span class="_"> </span>and<span class="_ _6"> </span>opti-</div><div class="t m0 x13 ha y34 ff2 fs4 fc0 sc0 ls0 ws0">mization<span class="_ _7"> </span>for<span class="_ _7"> </span>parallel<span class="_ _8"> </span>computations,<span class="_ _8"> </span>rather<span class="_ _7"> </span>than<span class="_ _7"> </span>the<span class="_ _8"> </span>low<span class="_ _7"> </span>com-</div><div class="t m0 x13 ha y35 ff2 fs4 fc0 sc0 ls0 ws0">putation<span class="_ _e"> </span>volume<span class="_ _e"> </span>theoretical<span class="_ _5"> </span>indicator<span class="_ _e"> </span>(BFLOP).<span class="_ _5"> </span>W<span class="_ _a"></span>e<span class="_ _5"> </span>hope</div><div class="t m0 x13 ha y36 ff2 fs4 fc0 sc0 ls0 ws0">that<span class="_"> </span>the<span class="_"> </span>designed<span class="_"> </span>object<span class="_ _6"> </span>can<span class="_"> </span>be<span class="_"> </span>easily<span class="_ _6"> </span>trained<span class="_"> </span>and<span class="_"> </span>used.<span class="_ _3"> </span>For</div><div class="t m0 x13 ha y37 ff2 fs4 fc0 sc0 ls0 ws0">example,<span class="_"> </span>an<span class="_ _0"></span>yone<span class="_"> </span>who<span class="_"> </span>uses<span class="_"> </span>a<span class="_ _8"> </span>con<span class="_ _0"></span>ventional<span class="_"> </span>GPU<span class="_"> </span>to<span class="_ _8"> </span>train<span class="_"> </span>and</div><div class="t m0 x13 ha y38 ff2 fs4 fc0 sc0 ls0 ws0">test<span class="_"> </span>can<span class="_ _6"> </span>achiev<span class="_ _0"></span>e<span class="_ _6"> </span>real-time,<span class="_ _6"> </span>high<span class="_"> </span>quality<span class="_ _0"></span>,<span class="_ _6"> </span>and<span class="_"> </span>convincing<span class="_"> </span>ob-</div><div class="t m0 x13 ha y39 ff2 fs4 fc0 sc0 ls0 ws0">ject<span class="_"> </span>detection<span class="_"> </span>results,<span class="_"> </span>as<span class="_"> </span>the<span class="_ _6"> </span>Y<span class="_ _0"></span>OLOv4<span class="_ _6"> </span>results<span class="_"> </span>shown<span class="_"> </span>in<span class="_"> </span>Fig-</div><div class="t m0 x13 ha y3a ff2 fs4 fc0 sc0 ls0 ws0">ure<span class="_"> </span><span class="fc2">1</span>.<span class="_ _4"> </span>Our<span class="_"> </span>contributions<span class="_"> </span>are<span class="_"> </span>summarized<span class="_"> </span>as<span class="_"> </span>follo<span class="_ _0"></span>ws:</div><div class="t m0 x15 ha y3b ff2 fs4 fc0 sc0 ls0 ws0">1.<span class="_ _13"> </span>W<span class="_ _a"></span>e<span class="_"> </span>de<span class="_ _0"></span>velope<span class="_ _7"> </span>an<span class="_"> </span>ef<span class="_ _0"></span>&#64257;cient<span class="_ _8"> </span>and<span class="_ _8"> </span>powerful<span class="_ _7"> </span>object<span class="_ _8"> </span>detection</div><div class="t m0 x16 ha y3c ff2 fs4 fc0 sc0 ls0 ws0">model.<span class="_ _6"> </span>It<span class="_"> </span>makes<span class="_ _7"> </span>ev<span class="_ _0"></span>eryone<span class="_"> </span>can<span class="_ _7"> </span>use<span class="_"> </span>a<span class="_ _7"> </span>1080<span class="_"> </span>T<span class="_ _0"></span>i<span class="_ _8"> </span>or<span class="_ _8"> </span>2080<span class="_ _8"> </span>T<span class="_ _0"></span>i</div><div class="t m0 x16 ha y3d ff2 fs4 fc0 sc0 ls0 ws0">GPU<span class="_"> </span>to<span class="_"> </span>train<span class="_"> </span>a<span class="_"> </span>super<span class="_"> </span>fast<span class="_"> </span>and<span class="_"> </span>accurate<span class="_"> </span>object<span class="_"> </span>detector<span class="_ _0"></span>.</div><div class="t m0 x15 ha y3e ff2 fs4 fc0 sc0 ls0 ws0">2.<span class="_ _13"> </span>W<span class="_ _a"></span>e<span class="_ _9"> </span>v<span class="_ _0"></span>erify<span class="_ _14"> </span>the<span class="_ _11"> </span>in&#64258;uence<span class="_ _14"> </span>of<span class="_ _11"> </span>state-of-the-art<span class="_ _14"> </span>Bag-of-</div><div class="t m0 x16 ha y3f ff2 fs4 fc0 sc0 ls0 ws0">Freebies<span class="_ _8"> </span>and<span class="_"> </span>Bag-of-Specials<span class="_ _7"> </span>methods<span class="_"> </span>of<span class="_ _7"> </span>object<span class="_"> </span>detec-</div><div class="t m0 x16 ha y40 ff2 fs4 fc0 sc0 ls0 ws0">tion<span class="_"> </span>during<span class="_"> </span>the<span class="_"> </span>detector<span class="_"> </span>training.</div><div class="t m0 x15 ha y41 ff2 fs4 fc0 sc0 ls0 ws0">3.<span class="_ _13"> </span>W<span class="_ _a"></span>e<span class="_ _1"> </span>modify<span class="_ _5"> </span>state-of-the-art<span class="_ _1"> </span>methods<span class="_ _5"> </span>and<span class="_ _5"> </span>make<span class="_ _1"> </span>them</div><div class="t m0 x16 ha y42 ff2 fs4 fc0 sc0 ls0 ws0">more<span class="_ _e"> </span>effecient<span class="_ _3"> </span>and<span class="_ _5"> </span>suitable<span class="_ _e"> </span>for<span class="_ _e"> </span>single<span class="_ _e"> </span>GPU<span class="_ _e"> </span>training,</div><div class="t m0 x16 ha y43 ff2 fs4 fc0 sc0 ls0 ws0">including<span class="_"> </span>CBN<span class="_"> </span>[<span class="fc3">89</span>],<span class="_"> </span>P<span class="_ _a"></span>AN<span class="_"> </span>[<span class="fc3">49</span>],<span class="_"> </span>SAM<span class="_"> </span>[<span class="fc3">85</span>],<span class="_"> </span>etc.</div><div class="t m0 x17 ha y44 ff2 fs4 fc0 sc0 ls0 ws0">1</div><div class="t m1 x18 hb y45 ff9 fs6 fc4 sc0 ls0 ws0">arXiv:2004.10934v1 [cs.CV] 23 Apr 2020</div><a class="l" rel='nofollow' onclick='return false;'><div class="d m2"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m2"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m2"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m2"></div></a><a class="l" rel='nofollow' onclick='return false;'><div class="d m2"></div></a></div><div class="pi" data-data='{"ctm":[1.568627,0.000000,0.000000,1.568627,0.000000,0.000000]}'></div></div> </body> </html>
评论
    相关推荐