%PDF-1.3 1 0 obj << /Kids [ 4 0 R 5 0 R 6 0 R 7 0 R 8 0 R 9 0 R 10 0 R 11 0 R 12 0 R 13 0 R 14 0 R ] /Type /Pages /Count 11 >> endobj 2 0 obj << /Subject (Neural Information Processing Systems http\072\057\057nips\056cc\057) /Publisher (Curran Associates\054 Inc\056) /Language (en\055US) /Created (2019) /EventType (Poster) /Description-Abstract (Object detectors are usually equipped with backbone networks designed for image classification\056 It might be sub\055optimal because of the gap between the tasks of image classification and object detection\056 In this work\054 we present DetNAS to use Neural Architecture Search \050NAS\051 for the design of better backbones for object detection\056 It is non\055trivial because detection training typically needs ImageNetpre\055training while NAS systems require accuracies on the target detection task as supervisory signals\056 Based on the technique of one\055shot supernet\054 which contains all possible networks in the search space\054 we propose a framework for backbone search on object detection\056 We train the supernet under the typical detector training schedule\072 ImageNet pre\055training and detection fine\055tuning\056 Then\054 the architecture search is performed on the trained supernet\054 using the detection task as the guidance\056 This framework makes NAS on backbones very efficient\056 In experiments\054 we show the effectiveness of DetNAS on various detectors\054 for instance\054 one\055stage RetinaNetand the two\055stage FPN\056 We empirically find that networks searched on object detection shows consistent superiority compared to those searched on ImageNet classification\056 The resulting architecture achieves superior performance than hand\055crafted networks on COCO with much less FLOPs complexity\056) /Producer (PyPDF2) /Title (DetNAS\072 Backbone Search for Object Detection) /Date (2019) /ModDate (D\07220200213010214\05508\04700\047) /Published (2019) /Type (Conference Proceedings) /firstpage (6642) /Book (Advances in Neural Information Processing Systems 32) /Description (Paper accepted and presented at the Neural Information Processing Systems Conference \050http\072\057\057nips\056cc\057\051) /Editors (H\056 Wallach and H\056 Larochelle and A\056 Beygelzimer and F\056 d\047Alch\351\055Buc and E\056 Fox and R\056 Garnett) /Author (Yukang Chen\054 Tong Yang\054 Xiangyu Zhang\054 GAOFENG MENG\054 Xinyu Xiao\054 Jian Sun) /lastpage (6652) >> endobj 3 0 obj << /Type /Catalog /Pages 1 0 R >> endobj 4 0 obj << /Parent 1 0 R /Contents 15 0 R /Resources 16 0 R /Rotate 0 /MediaBox [ 0 0 612 792 ] /Annots 77 0 R /Type /Page >> endobj 5 0 obj << /Parent 1 0 R /Contents 92 0 R /Resources 93 0 R /Rotate 0 /MediaBox [ 0 0 612 792 ] /Annots 129 0 R /Type /Page >> endobj 6 0 obj << /Parent 1 0 R /Contents 139 0 R /Resources 140 0 R /Rotate 0 /MediaBox [ 0 0 612 792 ] /Annots 166 0 R /Type /Page >> endobj 7 0 obj << /Parent 1 0 R /Contents 193 0 R /Resources 194 0 R /Rotate 0 /MediaBox [ 0 0 612 792 ] /Annots 214 0 R /Type /Page >> endobj 8 0 obj << /Parent 1 0 R /Contents 228 0 R /Resources 229 0 R /Rotate 0 /MediaBox [ 0 0 612 792 ] /Annots 249 0 R /Type /Page >> endobj 9 0 obj << /Parent 1 0 R /Contents 259 0 R /Resources 260 0 R /Rotate 0 /MediaBox [ 0 0 612 792 ] /Annots 261 0 R /Type /Page >> endobj 10 0 obj << /Parent 1 0 R /Contents 267 0 R /Resources 268 0 R /Rotate 0 /MediaBox [ 0 0 612 792 ] /Annots 269 0 R /Type /Page >> endobj 11 0 obj << /Parent 1 0 R /Contents 278 0 R /Resources 279 0 R /Rotate 0 /MediaBox [ 0 0 612 792 ] /Annots 286 0 R /Type /Page >> endobj 12 0 obj << /Parent 1 0 R /Contents 297 0 R /Resources 298 0 R /Rotate 0 /MediaBox [ 0 0 612 792 ] /Annots 394 0 R /Type /Page >> endobj 13 0 obj << /Parent 1 0 R /Contents 396 0 R /Resources 397 0 R /Rotate 0 /MediaBox [ 0 0 612 792 ] /Type /Page >> endobj 14 0 obj << /Parent 1 0 R /Contents 398 0 R /Resources 399 0 R /Rotate 0 /MediaBox [ 0 0 612 792 ] /Type /Page >> endobj 15 0 obj << /Length 4829 /Filter /FlateDecode >> stream x[rȕ}^@?Yv)p`dUPheL1sL >03Yソxy{2aQ%e%AYd|x}{E T^( $|Ye,nދȋ{'n/8o.oջuuWkv3まzbw/xaP叠,R<+,_Bo@Sy334^'AT&ޘ_,/=Xdu@̓r"0 pY]x7a Y(yJ"^#PTFBR3ԖiE\#_xP_`e0K{fABё h-N +|S=Pp|gaQii["մ^Z$y::4%q$={҄o)lШޜם,gŔa v[VW)Zi)V/:5[I~gu'͊ !A>e=[ޞdh-J.`s_Q?1nM:GqyAW3a\U?>Wm[?+$î!DQvimmR]oպ'_zGŘUSV/H LsV/'h\ Ry[2%Rh^Gn=k~.<?Qi7w磄`pb'ʿ~ KAS*XQp?A?KFB VXeQ3F (Jᾒ02/oOB<_1b&CQE{Y1µ@ǖ^ODzל@WU0?f ii5 ~ҁHۏS ܛmt* (J ˥զ:+ f習Ħ\zyXɭrhÆq^Ug-Na0&ڵ{ ) н~n{[rA#aIBT.٩(A~ 5 W;kQya y8섛L;5@.`5+{&p2FRYZ]t]* Z7LSϪyLu5原/˷?_<]FH/b