%PDF-1.3 1 0 obj << /Kids [ 3 0 R 4 0 R 5 0 R 6 0 R 7 0 R 8 0 R 9 0 R 10 0 R 11 0 R ] /Type /Pages /Count 9 >> endobj 2 0 obj << /Title (Learning Deep Features for Discriminative Localization) /Producer (PyPDF2) /Author (Bolei Zhou\054 Aditya Khosla\054 Agata Lapedriza\054 Aude Oliva\054 Antonio Torralba) /Subject (2016 IEEE Conference on Computer Vision and Pattern Recognition) >> endobj 3 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 13 0 R /Resources << /XObject << /x8 14 0 R /R36 17 0 R /R37 18 0 R /x6 19 0 R /x12 22 0 R /x10 25 0 R >> /ExtGState << /s9 28 0 R /s11 31 0 R /a0 << /CA 1 /ca 1 >> /R25 34 0 R /s5 35 0 R /s7 38 0 R >> /Font << /F2 41 0 R /R28 42 0 R /F1 46 0 R /R38 47 0 R /R34 51 0 R /R32 55 0 R /R30 58 0 R /R26 62 0 R >> /ProcSet [ /Text /ImageC /ImageB /PDF /ImageI ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 66 0 R 67 0 R 68 0 R 69 0 R 70 0 R 71 0 R 72 0 R 73 0 R 74 0 R 75 0 R 76 0 R 77 0 R 78 0 R 79 0 R 80 0 R 81 0 R 82 0 R ] >> endobj 4 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 83 0 R /Resources << /ExtGState << /R25 34 0 R >> /Font << /R79 84 0 R /R77 87 0 R /R75 91 0 R /F2 95 0 R /R28 42 0 R /F1 96 0 R /R83 97 0 R /R34 51 0 R /R87 100 0 R /R81 104 0 R /R85 108 0 R /R26 62 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 111 0 R 112 0 R 113 0 R 114 0 R 115 0 R 116 0 R 117 0 R 118 0 R 119 0 R 120 0 R 121 0 R 122 0 R 123 0 R 124 0 R 125 0 R 126 0 R 127 0 R 128 0 R 129 0 R 130 0 R 131 0 R 132 0 R 133 0 R 134 0 R 135 0 R 136 0 R 137 0 R ] >> endobj 5 0 obj << /Parent 1 0 R /Rotate 0 /Contents 138 0 R /Resources << /XObject << /R118 139 0 R /R119 140 0 R /R111 143 0 R /R116 144 0 R /R125 145 0 R /R114 150 0 R /R115 151 0 R /R131 152 0 R /R132 154 0 R /R133 155 0 R /R134 156 0 R /R124 147 0 R /R123 148 0 R /R128 158 0 R >> /ExtGState << /R103 163 0 R /R102 164 0 R /R104 165 0 R /R127 157 0 R /R122 146 0 R /R121 166 0 R /R120 167 0 R /R25 34 0 R >> /Font << /R109 168 0 R /R79 84 0 R /R112 172 0 R /R77 87 0 R /R107 175 0 R /R75 91 0 R /F2 179 0 R /R28 42 0 R /F1 180 0 R /R34 51 0 R /R81 104 0 R /R129 159 0 R /R26 62 0 R >> /ProcSet [ /Text /ImageC /ImageB /PDF /ImageI ] /ColorSpace << /R101 181 0 R /R106 141 0 R >> >> /Group 149 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 183 0 R 184 0 R 185 0 R 186 0 R 187 0 R 188 0 R ] /Type /Page >> endobj 6 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 189 0 R /Resources << /ExtGState << /R25 34 0 R >> /Font << /F2 190 0 R /R28 42 0 R /R79 84 0 R /F1 191 0 R /R34 51 0 R /R32 55 0 R /R30 58 0 R /R26 62 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 192 0 R 193 0 R 194 0 R 195 0 R 196 0 R 197 0 R 198 0 R 199 0 R 200 0 R 201 0 R 202 0 R 203 0 R 204 0 R 205 0 R 206 0 R 207 0 R 208 0 R 209 0 R 210 0 R ] >> endobj 7 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 211 0 R /Resources << /ExtGState << /R25 34 0 R >> /Font << /F2 212 0 R /R28 42 0 R /R79 84 0 R /R197 213 0 R /F1 217 0 R /R34 51 0 R /R32 55 0 R /R30 58 0 R /R26 62 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 218 0 R 219 0 R 220 0 R 221 0 R 222 0 R 223 0 R 224 0 R 225 0 R 226 0 R 227 0 R 228 0 R 229 0 R 230 0 R 231 0 R 232 0 R 233 0 R 234 0 R 235 0 R 236 0 R 237 0 R 238 0 R 239 0 R 240 0 R 241 0 R 242 0 R 243 0 R 244 0 R 245 0 R 246 0 R 247 0 R ] >> endobj 8 0 obj << /Parent 1 0 R /Rotate 0 /Contents 248 0 R /Resources << /XObject << /R228 249 0 R /R229 250 0 R /R223 251 0 R /R215 253 0 R /R216 254 0 R /R237 255 0 R /R236 256 0 R /R235 257 0 R /R234 259 0 R /R224 258 0 R /R232 262 0 R /R231 266 0 R /R230 260 0 R /R225 263 0 R /R226 264 0 R /R233 261 0 R /R227 265 0 R /R240 271 0 R /R219 276 0 R /R218 278 0 R /R222 252 0 R >> /ExtGState << /R25 34 0 R /R122 146 0 R /R217 277 0 R >> /Font << /R213 279 0 R /R238 272 0 R /R220 267 0 R /F2 283 0 R /R28 42 0 R /F1 284 0 R /R26 62 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /Group 149 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 285 0 R 286 0 R 287 0 R 288 0 R 289 0 R 290 0 R 291 0 R 292 0 R 293 0 R 294 0 R ] /Type /Page >> endobj 9 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 295 0 R /Resources << /XObject << /R264 296 0 R /R265 297 0 R /R266 298 0 R /R267 299 0 R /R261 300 0 R /R262 301 0 R /R263 302 0 R /R268 303 0 R /R269 304 0 R /R256 305 0 R /R258 306 0 R /R271 307 0 R /R270 308 0 R /R257 309 0 R >> /ExtGState << /R25 34 0 R >> /Font << /R79 84 0 R /F2 310 0 R /R28 42 0 R /F1 311 0 R /R259 312 0 R /R272 316 0 R /R34 51 0 R /R274 320 0 R /R32 55 0 R /R276 324 0 R /R30 58 0 R /R26 62 0 R >> /ProcSet [ /Text /ImageC /ImageB /PDF /ImageI ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 327 0 R 328 0 R 329 0 R 330 0 R 331 0 R 332 0 R 333 0 R 334 0 R 335 0 R ] >> endobj 10 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 336 0 R /Resources << /XObject << /R310 337 0 R /R309 338 0 R /R308 339 0 R /R303 340 0 R /R305 341 0 R /R304 342 0 R /R298 343 0 R /R288 344 0 R /R289 345 0 R /R291 346 0 R /R290 347 0 R /R293 348 0 R /R292 349 0 R /R295 350 0 R /R294 351 0 R /R297 352 0 R /R296 353 0 R >> /ExtGState << /R25 34 0 R >> /Font << /R299 354 0 R /R28 42 0 R /F1 358 0 R /F2 359 0 R /R34 51 0 R /R286 360 0 R /R306 364 0 R /R301 368 0 R /R26 62 0 R >> /ProcSet [ /Text /ImageC /ImageB /PDF /ImageI ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 372 0 R 373 0 R 374 0 R ] >> endobj 11 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 375 0 R /Resources << /ExtGState << /R25 34 0 R >> /Font << /F2 376 0 R /R28 42 0 R /R34 51 0 R /F1 377 0 R /R26 62 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 378 0 R 379 0 R 380 0 R 381 0 R 382 0 R 383 0 R 384 0 R 385 0 R 386 0 R 387 0 R 388 0 R 389 0 R 390 0 R 391 0 R 392 0 R 393 0 R 394 0 R 395 0 R 396 0 R 397 0 R 398 0 R 399 0 R 400 0 R 401 0 R 402 0 R 403 0 R 404 0 R 405 0 R 406 0 R 407 0 R 408 0 R 409 0 R 410 0 R 411 0 R 412 0 R 413 0 R 414 0 R 415 0 R 416 0 R 417 0 R 418 0 R 419 0 R 420 0 R 421 0 R 422 0 R 423 0 R 424 0 R 425 0 R 426 0 R 427 0 R 428 0 R 429 0 R 430 0 R 431 0 R 432 0 R 433 0 R 434 0 R 435 0 R ] >> endobj 12 0 obj << /Type /Catalog /Pages 1 0 R >> endobj 13 0 obj << /Length 19303 >> stream q q q 0.10000 0 0 0.10000 0 0 cm /R25 gs 0 g q 10 0 0 10 0 0 cm BT /R26 14.34620 Tf 1 0 0 1 126.81500 675.06700 Tm [ (Lear) 14.98930 (ning) -249.99400 (Deep) -250.00800 (F) 24.98890 (eatur) 17.99120 (es) -250.01200 (f) 24.99230 (or) -249.99500 (Discriminati) 10 (v) 9.99625 (e) -250 (Localization) ] TJ /R28 11.95520 Tf -11.17700 -37.85820 Td [ (Bolei) -249.99800 (Zhou\054) -249.98400 (Aditya) -250 (Khosla\054) -250.01600 (Ag) 4.98892 (ata) -249.99500 (Lapedriza\054) -250.01000 (Aude) -249.98500 (Oli) 25.01400 (v) 24.98340 (a\054) -249.98700 (Antonio) -250.01400 (T) 80.01470 (orralba) ] TJ 33.45000 -13.94690 Td [ (Computer) -250.01400 (Science) -250.00600 (and) -249.98700 <41727469026369616c> -249.99500 (Intelligence) -250 (Laboratory) 65 (\054) -250.01000 (MIT) ] TJ /R30 11.95520 Tf -26.02070 -13.94800 Td (f) Tj /R32 11.95520 Tf 5.97695 0 Td (bzhou\054khosla\054agata\054oliva\054torralba) Tj /R30 11.95520 Tf 236.71200 0 Td (g) Tj /R32 11.95520 Tf 5.97813 0 Td (\100csail\056mit\056edu) Tj /R26 11.95520 Tf -225.73900 -41.04610 Td (Abstract) Tj /R34 9.96260 Tf -83.92770 -24.70900 Td [ (In) -261.01300 (this) -260.99000 (work\054) -264.01200 (we) -261.00700 (r) 37.01960 (e) 15.01220 (visit) -260.99200 (the) -261.00800 (global) -260.98100 (aver) 15.01960 (a) 10.00320 (g) 10.00320 (e) -261.02000 (pooling) -261.01500 (layer) ] TJ -11.95510 -11.95510 Td [ (pr) 44.98390 (oposed) -213.00100 (in) -213.01100 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R34 9.96260 Tf 1 0 0 1 102.61500 531.60400 Tm (13) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R34 9.96260 Tf 1 0 0 1 112.57800 531.60400 Tm [ (\135\054) -221.01300 (and) -213.00900 (shed) -213.99100 (light) -212.99100 (on) -213.01300 (how) -213.00100 (it) -213.98900 (e) 19.99180 (xpl) 0.98758 (icitly) -213.98400 (enables) ] TJ -62.46560 -11.95510 Td [ (the) -384.98700 (con) 39.99880 (volutional) -386.00400 (neur) 14.99010 (al) -384.98200 (network) -385.00600 (\050CNN\051) -385 (to) -384.98200 (have) -386.01200 (r) 37.01830 (emark\055) ] TJ 11.95590 TL T* [ (able) -431.99300 (localization) -433.00300 (ability) -432.01100 (despite) -431.98400 (being) -433.00900 (tr) 14.99010 (ained) -431.98900 (on) -433.01600 (ima) 10.01300 (g) 10.00320 (e\055) ] TJ 11.95510 TL T* [ (le) 14.98100 (vel) -293.99100 (labels\056) -440.99600 (While) -294.01200 (this) -292.98400 (tec) 15.01590 (hnique) -294.00600 (was) -294.00500 (pr) 36.98650 (e) 15.01220 (viously) -294.00700 (pr) 44.98510 (oposed) ] TJ T* [ (as) -348.00900 (a) -348.99600 (means) -348.00600 (for) -348.98200 (r) 37.01960 (e) 39.98840 (gularizing) -348 (tr) 14.99140 (aining) 9.99343 (\054) -372.99400 (we) -348.01100 <026e64> -348.98600 (that) -347.99600 (it) -348.98600 (actu\055) ] TJ T* [ (ally) -387.01800 (b) 20.00160 (uilds) -387.00300 (a) -386.00900 (g) 10.00320 (eneric) -386.98900 (localizable) -386.98700 (deep) -386.99200 (r) 37.01830 (epr) 36.98160 (esentation) -386.99400 (that) ] TJ T* [ (e) 19.99240 (xposes) -201.99200 (the) -201.01700 (implicit) -202.00600 (attention) -202 (of) -201.99300 (CNNs) -201.01900 (on) -201.99600 (an) -201.99600 (ima) 10.01300 (g) 10.00320 (e) 15.01220 (\056) -293.98800 (Despite) ] TJ T* [ (the) -348.01200 (appar) 37 (ent) -346.99300 (simplicity) -348.01800 (of) -348.00800 (global) -347.00600 (aver) 15.01960 (a) 10.00320 (g) 10.00320 (e) -347.98600 (pooling) 9.99833 (\054) -372.01400 (we) -348.01100 (ar) 36.98650 (e) ] TJ 11.95590 TL T* [ (able) -233.00700 (to) -233.00800 (ac) 15.01770 (hie) 14.98590 (ve) -232.98900 (37\0561\045) -231.98100 (top\0555) -233.00200 (err) 44.98020 (or) -233.00900 (for) -233.00200 (object) -233.00500 (localization) -232.99500 (on) ] TJ 11.95510 TL T* [ (ILSVRC) -277.99500 (2014) -278.99300 (without) -277.98200 (tr) 14.99140 (aining) -279.01700 (on) -277.98300 (any) -279.00600 (bounding) -277.99000 (box) -279.00700 (anno\055) ] TJ T* [ (tation\056W) 91.98650 (e) -303.01300 (demonstr) 15.01100 (ate) -303 (in) -304.01500 (a) -303.00300 (variety) -302.99600 (of) -302.99600 (e) 19.99180 (xperiments) -302.98100 (that) -304.00300 (our) ] TJ T* [ (network) -240.01100 (is) -241.00500 (able) -239.98700 (to) -239.98700 (localize) -240.98900 (the) -239.99200 (discriminative) -239.98400 (ima) 10.01300 (g) 10.00320 (e) -240.98400 (r) 37.01830 (e) 39.98840 (gions) ] TJ T* [ (despite) -250.01500 (just) -250.01100 (being) -250.01800 (tr) 14.99140 (ained) -250.01800 (for) -250.01800 (solving) -250.00600 <636c6173736902636174696f6e> -250.02000 (task) ] TJ ET Q 1 0 0 rg q 10 0 0 10 0 0 cm BT /R28 6.97380 Tf 1 0 0 1 269.97500 391.75700 Tm (1) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R34 9.96260 Tf 1 0 0 1 273.96000 388.14200 Tm (\056) Tj /R26 11.95520 Tf -223.84800 -38.26090 Td [ (1\056) -249.99000 (Intr) 18.01460 (oduction) ] TJ /R28 9.96260 Tf 11.95470 -19.72810 Td [ (Recent) -271.98200 (w) 10.00140 (ork) -273.00600 (by) -271.98400 (Zhou) ] TJ /R34 9.96260 Tf 89.88630 0 Td [ (et) -271.99100 (al) ] TJ /R28 9.96260 Tf 20.37970 0 Td (\133) Tj ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R28 9.96260 Tf 1 0 0 1 175.65000 330.15300 Tm (34) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R28 9.96260 Tf 1 0 0 1 185.61300 330.15300 Tm [ (\135) -271.99700 (has) -273.00600 (sho) 24.99340 (wn) -271.98600 (that) -272.01800 (the) -273.00600 (con\055) ] TJ -135.50100 -11.95510 Td [ (v) 20.00160 (olutional) -402.98700 (units) -403 (of) -402.99000 (v) 24.98110 (arious) -403.01500 (layers) -402.98500 (of) -402.99000 (con) 39.99820 (v) 20.00160 (olutional) -402.98700 (neural) ] TJ 11.95510 TL T* [ (netw) 10.00870 (orks) -395.01900 (\050CNNs\051) -395.01400 (actually) -395.01400 (beha) 19.99790 (v) 14.98280 (e) -394.99800 (as) -394.99200 (object) -395.01700 (detectors) -395.01700 (de\055) ] TJ T* [ (spite) -306.00700 (no) -306.01800 (supervision) -305.98300 (on) -306.01800 (the) -306.02000 (location) -306 (of) -305.98600 (the) -306.02000 (object) -306.01300 (w) 10 (as) -305.98600 (pro\055) ] TJ 11.95630 TL T* [ (vided\056) -738.99400 (Despite) -393.01300 (ha) 19.99790 (ving) -393.01000 (this) -393.98700 (rem) 0.98513 (arkable) -394.01300 (ability) -392.99800 (to) -392.98100 (localize) ] TJ 11.95510 TL T* [ (objects) -280.01000 (in) -279 (the) -279.98500 (con) 40 (v) 20.00160 (olutional) -279.00700 (layers\054) -287.00200 (this) -280.00700 (ability) -279.99700 (is) -279 (lost) -280.00700 (when) ] TJ T* [ (fully\055connected) -280.00400 (layers) -281.00600 (are) -280.00500 (used) -279.98000 (for) -281.01300 <636c6173736902636174696f6e2e> -400.98000 (Recently) ] TJ T* [ (some) -344.00200 (popular) -344.00100 (fully\055con) 40.01660 (v) 20.00160 (olutional) -344.01600 (neural) -344.01600 (netw) 10.00810 (orks) -344.98900 (such) -344.00900 (as) ] TJ T* [ (the) -348.99300 (Netw) 10.00690 (ork) -350.01300 (in) -348.98800 (Netw) 10.00690 (ork) -350.01300 (\050NIN\051) -349.00400 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R28 9.96260 Tf 1 0 0 1 184.61700 234.51200 Tm (13) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R28 9.96260 Tf 1 0 0 1 194.58000 234.51200 Tm [ (\135) -349.98400 (and) -348.99600 (GoogLeNet) -349.99800 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R28 9.96260 Tf 1 0 0 1 273.08500 234.51200 Tm (25) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R28 9.96260 Tf 1 0 0 1 283.04700 234.51200 Tm (\135) Tj -232.93500 -11.95510 Td [ (ha) 19.99730 (v) 14.98280 (e) -236.00400 (been) -237.01800 (proposed) -236.01600 (to) -237.00700 (a) 19.99180 (v) 20.00160 (oid) -235.98300 (the) -235.99200 (use) -237.01200 (of) -235.99900 (fully\055connected) -236.99200 (lay\055) ] TJ 11.95590 TL T* [ (ers) -305.00900 (to) -304.99600 (minimize) -305.00500 (the) -305 (number) -304.98800 (of) -305.00700 (parameters) -305 (while) -304.99100 (maintain\055) ] TJ 11.95510 TL T* [ (ing) -250.02000 (high) -250.01500 (performance\056) ] TJ 11.95510 -12.75310 Td [ (In) -313.98500 (order) -313.98700 (to) -314.01400 (achie) 25.01420 (v) 14.98280 (e) -313.99200 (this\054) -329.99500 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R28 9.96260 Tf 1 0 0 1 165.17000 185.89300 Tm (13) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R28 9.96260 Tf 1 0 0 1 175.13200 185.89300 Tm [ (\135) -313.99100 (uses) ] TJ /R34 9.96260 Tf 26.73550 0 Td [ (global) -313.99200 (aver) 15.01960 (a) 10.00320 (g) 10.00320 (e) -313.99200 (pool\055) ] TJ -151.75600 -11.95590 Td (ing) Tj /R28 9.96260 Tf 15.89880 0 Td [ (which) -318.00900 (acts) -317.98100 (as) -317.98500 (a) -317.01100 (structural) -318.00300 (re) 15.00980 (gularizer) 39.99820 (\054) -335 (pre) 25.01300 (v) 14.98280 (enting) -318 (o) 14.98280 (v) 14.98280 (er) 19.98930 (\055) ] TJ -15.89880 -11.95510 Td [ <027474696e67> -348.01400 (during) -347.99600 (training\056) -603.98500 (In) -348.01900 (our) -346.99400 (e) 15.01220 (xperiments\054) -372.99700 (we) -347.98400 (found) -348.00300 (that) ] TJ 11.95510 TL T* [ (the) -314.01800 (adv) 24.98170 (antages) -314.99600 (of) -313.98600 (this) -314.98100 (global) -314.00200 (a) 19.99180 (v) 14.98280 (erage) -315.01700 (pooling) -313.98700 (layer) -315.00600 (e) 15.01220 (xtend) ] TJ T* [ (be) 15.01770 (yond) -278.01200 (simply) -277.98900 (acting) -278.01600 (as) -277.99200 (a) -277.01700 (re) 15.00980 (gularizer) -277.98800 (\055) -277.99500 (In) -277.99000 (f) 9.99588 (act\054) -285.01100 (with) -278.01000 (a) -277.99800 (little) ] TJ T* [ (tweaking\054) -224.99600 (the) -217.99500 (netw) 10.00940 (ork) -217.99500 (can) -218.98700 (retain) -217.99500 (its) -217.98200 (remarkable) -219.01700 (localization) ] TJ T* [ (ability) -202.01000 (until) -202.99400 (the) -201.99800 <026e616c> -201.98300 (layer) 55.01040 (\056) -295.00700 (This) -201.98300 (tweaking) -201.98100 (allo) 24.99090 (ws) -203.01500 (identifying) ] TJ 11.95590 TL T* [ (easily) -251 (the) -249.99000 (discriminati) 24.98970 (v) 14.98280 (e) -250.98300 (image) -249.98200 (re) 15.00980 (gions) -250.98800 (in) -251.00500 (a) -250.00200 (single) -250.99000 (forw) 10 (ard\055) ] TJ ET Q 3.98000 w 0 G 501.12100 906.78900 m 1446.11000 906.78900 l S q 10 0 0 10 0 0 cm BT /R28 5.97760 Tf 1 0 0 1 60.97110 83.81290 Tm (1) Tj /R28 7.97010 Tf 3.48672 -2.81289 Td [ (Code) -250.02200 (and) -249.97800 (models) -249.97800 (are) -249.97600 (a) 20.00270 (v) 24.98200 (ailable) -249.97600 (at) -249.99300 (http\072\057\057cnnlocalization\056csail\056mit\056edu) ] TJ ET Q q 3088.62000 4584.97000 2361.43000 1196.77000 re W* n 1 1 1 rg 3088.62000 4584.97000 2361.43000 1197.34000 re f* q 1088.96000 0 0 1118.78000 3112.13000 4584.97000 cm /R36 Do Q q 1123.94000 0 0 1116.49000 4289.98000 4586.11000 cm /R37 Do Q 0 0 0 rg q 10 0 0 10 0 0 cm BT /R38 8.02816 Tf 1 0 0 1 341.08900 572.15200 Tm [ <01> 2.97973 <02> -2.00101 <0304> -5.98479 <0506> 0.99747 <070809> -6.98834 <0a> 0.02382 <0b0b0a> -6.98834 <05> -8076.89000 <0c> -6.00304 <030a> -6.98226 <0a> 0.02433 <06> 1.00355 <070809> -6.98226 <0a> 0.02433 <02> -2.00101 <0b0b04> ] TJ ET Q Q q 10 0 0 10 0 0 cm BT /R28 8.96640 Tf 1 0 0 1 308.86200 448.79900 Tm [ (Figure) -533.98700 (1\056) -533.00100 (A) -533.99200 (simple) -533.01200 <6d6f646902636174696f6e> -534.00900 (of) -533.99800 (the) -532.99000 (global) -533.98100 (a) 20.02380 (v) 14.98650 (erage) -533.00700 (pool\055) ] TJ 10.95900 TL T* [ (ing) -467.01300 (layer) -467.02100 (combined) -465.99800 (with) -467.01300 (our) -466.99400 (class) -467 (acti) 24.98480 (v) 25.00660 (ation) -466.97800 (mapping) -466.01900 (\050CAM\051) ] TJ T* [ (technique) -371 (allo) 25.00660 (ws) -371.00900 (the) -372.01600 <636c6173736902636174696f6e2d747261696e6564> -371.01900 (CNN) -371.00900 (to) -370.99200 (both) -370.99800 (classify) ] TJ T* [ (the) -423.99000 (image) -422.99600 (and) -423.99800 (localize) -422.98200 <636c6173732d73706563690263> -423.99500 (image) -424 (re) 14.99740 (gions) -423.00900 (in) -424.01100 (a) -422.99300 (single) ] TJ T* [ (forw) 9.98193 (ard\055pass) -234.00300 (e\056g\056\054) -237.99500 (the) -235.00300 (toothbrush) -233.98400 (for) ] TJ /R34 8.96640 Tf 132.21400 0 Td [ (brushing) -234.01700 (teeth) ] TJ /R28 8.96640 Tf 53.51990 0 Td [ (and) -234.01200 (the) -235.00300 (chain\055) ] TJ -185.73400 -10.95900 Td [ (sa) 15.01650 (w) -249.98900 (for) ] TJ /R34 8.96640 Tf 28.74610 0 Td [ (cutting) -250.01400 (tr) 37 (ees) ] TJ /R28 8.96640 Tf 44.24770 0 Td (\056) Tj /R28 9.96260 Tf -72.99380 -21.45390 Td [ (pass) -241.00600 (for) -240.98200 (a) -242.00400 (wide) -241.00900 (v) 24.98110 (ariety) -241.01100 (of) -241.01600 (tasks\054) -242.99400 (e) 25.01050 (v) 14.98280 (en) -241.01900 (those) -240.99900 (that) -241.98400 (the) -241.00900 (netw) 10.00810 (ork) ] TJ 11.95510 TL T* [ (w) 10 (as) -325.00200 (not) -324.00700 (originally) -325.00200 (trained) -324.01000 (for) 55.01040 (\056) -533.98400 (As) -324.99500 (sho) 24.99090 (wn) -324.01500 (in) -324.99500 (Figure) ] TJ ET Q 1 0 0 rg q 10 0 0 10 0 0 cm BT /R28 9.96260 Tf 1 0 0 1 518.73400 360.59500 Tm (1) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R28 9.96260 Tf 1 0 0 1 523.71600 360.59500 Tm [ (\050a\051\054) -343.99200 (a) ] TJ -214.85400 -11.95470 Td [ (CNN) -217.00300 (trained) -215.99100 (on) -217.01300 (object) -215.98800 (cate) 15.01220 (gorization) -216.99300 (is) -217.00800 (successfull) 1.00473 (y) -217.01800 (able) -217.01800 (to) ] TJ 11.95630 TL T* [ (localize) -322.99500 (the) -322.99800 (discriminati) 24.98850 (v) 14.98280 (e) -323.01000 (re) 15.00730 (gions) -323.01500 (for) -323.98500 (action) -322.99000 <636c6173736902636174696f6e> ] TJ 11.95470 TL T* [ (as) -376.01300 (the) -376.00800 (objects) -375.99400 (that) -376 (the) -376.00800 (humans) -375.98600 (are) -377.00600 (interacting) -375.98600 (with) -375.99100 (rather) ] TJ T* [ (than) -249.98500 (the) -249.99000 (humans) -250.00700 (themselv) 15.01220 (es\056) ] TJ 11.95510 -12.48200 Td [ (Despite) -288.99100 (the) -289.00300 (apparent) -289.98600 (simplicity) -288.99100 (of) -289.01100 (our) -289.00600 (approach\054) -299.00400 (for) -289.99100 (the) ] TJ -11.95510 -11.95510 Td [ (weakly) -360.00600 (supervised) -361.01100 (object) -360.00400 (localization) -359.98400 (on) -360.98900 (ILS) 0.99493 (VRC) -361.00800 (bench\055) ] TJ T* [ (mark) -326.01400 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R28 9.96260 Tf 1 0 0 1 335.90100 276.38200 Tm (21) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R28 9.96260 Tf 1 0 0 1 345.86300 276.38200 Tm [ (\135\054) -345.98100 (our) -326.01700 (best) -326.00900 (netw) 10.00810 (ork) -327 (achie) 25.01540 (v) 14.98280 (es) -325.98500 (37\0561\045) -326.00400 (top\0555) -326.00400 (test) -326.99000 (er) 19.98690 (\055) ] TJ -37.00120 -11.95470 Td [ (ror) 39.99090 (\054) -531.98700 (which) -475.98400 (is) -475.98700 (rather) -474.98200 (close) -475.98900 (to) -475.98900 (the) -475.98900 (34\0562\045) -475 (top\0555) -476.01900 (test) -475.98400 (error) ] TJ 11.95510 TL T* [ (achie) 25.01540 (v) 14.98280 (ed) -366.99800 (by) -366.98800 (fully) -367.01500 (supervised) -367.00700 (Ale) 14.98280 (xNet) -366.98800 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R28 9.96260 Tf 1 0 0 1 470.46600 252.47200 Tm (10) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R28 9.96260 Tf 1 0 0 1 480.42800 252.47200 Tm [ (\135\056) -660.98800 (Furthermore\054) ] TJ -171.56600 -11.95630 Td [ (we) -202.98800 (demonstrate) -204.01300 (that) -203.01000 (the) -203.01800 (localizability) -203.01500 (of) -204.00500 (the) -203.01500 (deep) -202.98600 (features) -204.00500 (in) ] TJ 11.95470 TL T* [ (our) -291.00300 (approach) -291.98600 (can) -291.01300 (be) -291.99000 (easily) -290.99600 (transferred) -291.99500 (to) -291 (other) -291.98100 (recognition) ] TJ T* [ (datasets) -248.99800 (for) -250 (generic) -249.01700 <636c6173736902636174696f6e2c> -249.98800 (localization\054) -248.98300 (and) -249.99300 (concept) ] TJ T* [ (disco) 14.99750 (v) 14.98280 (ery) 65.00630 (\056) ] TJ /R26 10.95890 Tf 21.50510 TL T* [ (1\0561\056) -250.00400 (Related) -249.99300 (W) 74.98940 (ork) ] TJ /R28 9.96260 Tf 11.95510 -18.45980 Td [ (Con) 40.01540 (v) 20.00160 (olutional) -263.01000 (Neural) -263.99300 (Netw) 10.00810 (orks) -263 (\050CNNs\051) -264.01500 (ha) 19.99670 (v) 14.98280 (e) -263.02000 (led) -263.00500 (to) -263.98500 (im\055) ] TJ -11.95510 -11.95510 Td [ (pressi) 24.99340 (v) 14.98280 (e) -492.98100 (performance) -493.01600 (on) -493.00600 (a) -492.98100 (v) 24.98110 (ariety) -493.00600 (of) -493.01600 (visual) -492.99100 (recognition) ] TJ 11.95510 TL T* [ (tasks) -244.01800 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R28 9.96260 Tf 1 0 0 1 334.53600 140.77600 Tm (10) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R28 9.96260 Tf 1 0 0 1 344.49800 140.77600 Tm (\054) Tj ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R28 9.96260 Tf 1 0 0 1 349.41000 140.77600 Tm [ (3) -1.01454 (5) ] TJ ET Q 0 g q 10 0 0 10 0 0 cm BT /R28 9.96260 Tf 1 0 0 1 359.38300 140.77600 Tm (\054) Tj ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R28 9.96260 Tf 1 0 0 1 364.29500 140.77600 Tm (8) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R28 9.96260 Tf 1 0 0 1 369.27600 140.77600 Tm [ (\135\056) -307.98800 (Recent) -243.98600 (w) 10.00320 (ork) -243.99400 (has) -243.01300 (sho) 24.99090 (wn) -243.98900 (that) -243.98400 (despite) -244.01800 (being) ] TJ -60.41370 -11.95510 Td [ (trained) -366.98500 (on) -366.98800 (image\055le) 24.99090 (v) 14.98280 (el) -368.01500 (labels\054) -395.99300 (CNNs) -367.01200 (ha) 19.99670 (v) 14.98280 (e) -367.98300 (the) -366.98800 (remarkable) ] TJ 11.95590 TL T* [ (ability) -224.98700 (to) -225.00900 (localize) -225.01100 (objects) -225 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R28 9.96260 Tf 1 0 0 1 413.56900 116.86500 Tm (1) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R28 9.96260 Tf 1 0 0 1 418.55000 116.86500 Tm (\054) Tj ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R28 9.96260 Tf 1 0 0 1 423.28300 116.86500 Tm (16) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R28 9.96260 Tf 1 0 0 1 433.24500 116.86500 Tm (\054) Tj ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R28 9.96260 Tf 1 0 0 1 437.97800 116.86500 Tm (2) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R28 9.96260 Tf 1 0 0 1 442.95900 116.86500 Tm (\054) Tj ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R28 9.96260 Tf 1 0 0 1 447.69100 116.86500 Tm (15) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R28 9.96260 Tf 1 0 0 1 457.65400 116.86500 Tm (\054) Tj ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R28 9.96260 Tf 1 0 0 1 462.38600 116.86500 Tm (18) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R28 9.96260 Tf 1 0 0 1 472.34900 116.86500 Tm [ (\135\056) -301.98900 (In) -225.02100 (this) -224.99700 (w) 10.00320 (ork\054) -229.99100 (we) ] TJ -163.48700 -11.95510 Td [ (sho) 24.99340 (w) -267.99200 (that\054) -272.99600 (using) -268.00400 (an) -267.99500 (appropriate) -269 (architecture\054) -271.98900 (we) -267.99500 (can) -269.01900 (gener) 19.99670 (\055) ] TJ 11.95510 TL T* [ (alize) -206.98300 (this) -206.02000 (ability) -206.99000 (be) 15.01710 (yond) -207.00500 (just) -206.02000 (localizing) -206.99500 (objects\054) -214.99800 (to) -207.01400 (start) -207.01000 (iden\055) ] TJ T* [ (tifying) -214.00800 (e) 15.01220 (xactly) -213.98900 (which) -213.00900 (re) 15.00980 (gions) -214.01300 (of) -214.00400 (an) -214.00400 (image) -213.00900 (are) -214.01300 (being) -213.98400 (used) -213.99400 (for) ] TJ -13.74100 -29.88790 Td (1) Tj ET Q Q Q q q 1 1 1 rg /a0 gs 48.40600 786.42200 515.18800 -52.69900 re f q /s5 gs /x6 Do Q q /s7 gs /x8 Do Q q /s9 gs /x10 Do Q q /s11 gs /x12 Do Q Q Q Q q 1 0 0 1 0 0 cm BT /F1 12 Tf 14.40000 TL ET 1 1 1 rg n 270 47 72 14 re f* 0.50000 0.50000 0.50000 rg BT /F2 9 Tf 10.80000 TL ET BT 1 0 0 1 297 50 Tm (2921) Tj T* ET Q endstream endobj 14 0 obj << /Filter /FlateDecode /Resources << /ExtGState << /a0 << /CA 1 /ca 1 >> >> /XObject << /x18 15 0 R >> >> /Length 28 /Group << /Type /Group /S /Transparency /CS /DeviceRGB /I true >> /BBox [ 78 746 96 765 ] /Type /XObject /Subtype /Form >> stream x+O4PH/VЯ0Pp 0 endstream endobj 15 0 obj << /Filter /FlateDecode /Resources 16 0 R /Length 107 /Type /XObject /BBox [ 78 746 96 765 ] /Subtype /Form >> stream xe AC̬wʠ =p,?]%+H-
Jc "82w8VSnGW;"
endstream
endobj
16 0 obj
<<
/ExtGState <<
/a0 <<
/CA 1
/ca 1
>>
>>
>>
endobj
17 0 obj
<<
/Filter /DCTDecode
/BitsPerComponent 8
/Height 301
/Length 12254
/ColorSpace /DeviceRGB
/Width 293
/Subtype /Image
>>
stream
Adobe d C
$, !$4.763.22:ASF:=N>22HbINVX]^]8EfmeZlS[]Y C**Y;2;YYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYY -%"
} !1AQa"q2#BR$3br
%&'()*456789:CDEFGHIJSTUVWXYZcdefghijstuvwxyz
w !1AQaq"2B #3Rbr
$4%&'()*56789:CDEFGHIJSTUVWXYZcdefghijstuvwxyz ? 赍RS(&
BSUKUO_5α/;OfΣG8G֡ ? SҐMIZ (:֢ Om!BΡ> /S 5 /U >aRbcgQ| C~?
R▣-cP Ӭ;A U+h_h,'ZԇY
pֵ / Taj+u *.xyTx7bi뺛u_[GKq³-ڥsRR.EjM