%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 (Where to Look\072 Focus Regions for Visual Question Answering) /Producer (PyPDF2) /Author (Kevin J\056 Shih\054 Saurabh Singh\054 Derek Hoiem) /Subject (2016 IEEE Conference on Computer Vision and Pattern Recognition) >> endobj 3 0 obj << /Parent 1 0 R /Rotate 0 /Contents 13 0 R /Resources << /XObject << /x8 14 0 R /R25 17 0 R /x6 18 0 R /R42 21 0 R /R41 38 0 R /R40 39 0 R /R24 37 0 R /R30 26 0 R /R37 31 0 R /R34 34 0 R /R35 33 0 R /x12 44 0 R /R23 32 0 R /x10 47 0 R >> /ExtGState << /s9 50 0 R /R39 29 0 R /s5 53 0 R /s7 56 0 R /R29 27 0 R /R11 59 0 R /s11 60 0 R /a0 << /CA 1 /ca 1 >> /R36 22 0 R /R22 23 0 R /R32 24 0 R >> /Font << /F2 63 0 R /R12 64 0 R /F1 67 0 R /R14 68 0 R /R16 72 0 R /R18 76 0 R /R20 79 0 R /R26 40 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /Group 36 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 83 0 R 84 0 R ] /Type /Page >> endobj 4 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 85 0 R /Resources << /XObject << /R23 32 0 R /R87 86 0 R /R84 87 0 R >> /ExtGState << /R11 59 0 R >> /Font << /F1 88 0 R /F2 89 0 R /R12 64 0 R /R88 90 0 R /R14 68 0 R /R85 95 0 R >> /ProcSet [ /Text /ImageC /ImageB /PDF /ImageI ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 99 0 R 100 0 R 101 0 R 102 0 R 103 0 R 104 0 R 105 0 R 106 0 R 107 0 R 108 0 R 109 0 R 110 0 R 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 ] >> endobj 5 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 133 0 R /Resources << /XObject << /R98 134 0 R /R99 135 0 R >> /ExtGState << /R11 59 0 R >> /Font << /R108 136 0 R /F1 139 0 R /R112 140 0 R /R100 144 0 R /R110 149 0 R /R102 152 0 R /R104 155 0 R /R106 160 0 R /F2 163 0 R /R12 64 0 R /R88 90 0 R /R14 68 0 R /R96 164 0 R >> /ProcSet [ /Text /ImageC /ImageB /PDF /ImageI ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 168 0 R ] >> endobj 6 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 169 0 R /Resources << /ExtGState << /R11 59 0 R >> /Font << /F2 170 0 R /R132 171 0 R /F1 175 0 R /R134 176 0 R /R100 144 0 R /R14 68 0 R /R102 152 0 R /R12 64 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 180 0 R 181 0 R 182 0 R 183 0 R 184 0 R 185 0 R 186 0 R 187 0 R 188 0 R 189 0 R 190 0 R 191 0 R ] >> endobj 7 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 192 0 R /Resources << /ExtGState << /R11 59 0 R >> /Font << /F2 193 0 R /R100 144 0 R /R12 64 0 R /F1 194 0 R /R14 68 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 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 8 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 211 0 R /Resources << /XObject << /R181 212 0 R /R180 213 0 R /R183 214 0 R /R182 215 0 R /R185 216 0 R /R184 217 0 R /R186 218 0 R /R174 219 0 R /R175 220 0 R /R176 221 0 R /R177 222 0 R /R173 223 0 R /R178 224 0 R /R179 225 0 R >> /ExtGState << /R11 59 0 R >> /Font << /F2 226 0 R /R12 64 0 R /R88 90 0 R /R14 68 0 R /R187 227 0 R /R100 144 0 R /F1 231 0 R /R102 152 0 R /R171 232 0 R >> /ProcSet [ /Text /ImageC /ImageB /PDF /ImageI ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 236 0 R 237 0 R 238 0 R 239 0 R 240 0 R 241 0 R 242 0 R ] >> endobj 9 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 243 0 R /Resources << /XObject << /R228 244 0 R /R229 247 0 R /R213 248 0 R /R223 249 0 R /R206 250 0 R /R207 251 0 R /R217 252 0 R /R205 253 0 R /R220 254 0 R /R221 255 0 R /R222 256 0 R /R209 257 0 R /R224 258 0 R /R225 259 0 R /R226 260 0 R /R227 261 0 R /R208 262 0 R /R212 263 0 R /R231 264 0 R /R233 265 0 R /R215 266 0 R /R230 267 0 R /R211 268 0 R /R214 269 0 R /R210 270 0 R /R216 271 0 R /R219 272 0 R /R232 273 0 R /R218 274 0 R /R198 275 0 R >> /ExtGState << /R11 59 0 R >> /Font << /R203 276 0 R /R201 280 0 R /R199 284 0 R /F2 288 0 R /F1 289 0 R /R14 68 0 R >> /ProcSet [ /Text /ImageC /ImageB /PDF /ImageI ] /ColorSpace << /R197 245 0 R /R195 290 0 R >> >> /MediaBox [ 0 0 612 792 ] /Annots [ ] >> endobj 10 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 292 0 R /Resources << /ExtGState << /R11 59 0 R >> /Font << /F2 293 0 R /R12 64 0 R /F1 294 0 R /R14 68 0 R /R187 227 0 R /R100 144 0 R /R102 152 0 R /R242 295 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 298 0 R 299 0 R ] >> endobj 11 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 300 0 R /Resources << /ExtGState << /R11 59 0 R >> /Font << /F2 301 0 R /R20 79 0 R /R12 64 0 R /F1 302 0 R /R14 68 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 303 0 R 304 0 R 305 0 R 306 0 R 307 0 R 308 0 R 309 0 R 310 0 R 311 0 R 312 0 R 313 0 R 314 0 R 315 0 R 316 0 R 317 0 R 318 0 R 319 0 R 320 0 R 321 0 R 322 0 R 323 0 R 324 0 R 325 0 R 326 0 R 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 12 0 obj << /Type /Catalog /Pages 1 0 R >> endobj 13 0 obj << /Length 13604 >> stream q q q 0.10000 0 0 0.10000 0 0 cm /R11 gs 0 g q 10 0 0 10 0 0 cm BT /R12 14.34620 Tf 1 0 0 1 105.31700 675.06700 Tm [ (Wher) 17.99800 (e) -249.99900 (T) 91.98970 (o) -250.00800 (Look\072) -310.01300 (F) 24.98890 (ocus) -250.01300 (Regions) -250.00200 (f) 24.99230 (or) -249.99500 (V) 37.01370 (isual) -250.00500 (Question) -249.99800 (Answering) ] TJ /R14 11.95520 Tf 76.70660 -47.82030 Td [ (K) 24.99670 (e) 24.98750 (vin) -250.00400 (J\056) -249.99500 (Shih\054) -250.00600 (Saurabh) -250.00200 (Singh\054) -249.98500 (and) -249.98700 (Derek) -249.99700 (Hoiem) ] TJ 11.23200 -23.91090 Td [ (Uni) 24.99570 (v) 14.98510 (ersity) -249.98900 (of) -250.01400 (Illinois) -250.00800 (at) -249.98700 (Urbana\055Champaign) ] TJ /R16 8.96640 Tf 10.98120 -20.92110 Td [ (f) -0.90126 ] TJ /R18 8.96640 Tf 4.60781 0 Td [ (kjshih2\054) -600.02100 (ss1\054) -600.00200 (dhoiem) ] TJ /R16 8.96640 Tf 107.59600 0 Td [ (g) -0.90126 ] TJ /R18 8.96640 Tf 4.60820 0 Td (\100illinois\056edu) Tj /R12 11.95520 Tf -175.05400 -49.06560 Td (Abstract) Tj /R20 9.96260 Tf -83.92770 -24.41600 Td [ (W) 91.98650 (e) -299.99300 (pr) 36.98520 (esent) -300.01900 (a) -299.98400 (method) -300.00800 (that) -300.00400 (learns) -300.00400 (to) -300.01600 (answer) -299.99700 (visual) -300.00100 (ques\055) ] TJ -11.95510 -11.95510 Td [ (tions) -325.99300 (by) -325.98400 (selecting) -326 (ima) 10.01300 (g) 10.00320 (e) -325.99000 (r) 37.01960 (e) 39.98840 (gions) -325.99600 (r) 37.01960 (ele) 15.01590 (vant) -326.01200 (to) -326.01200 (the) -326.01700 (te) 20 (xt\055based) ] TJ 11.95630 TL T* [ (query) 55.01780 (\056) -301.00600 (Our) -221.98800 (method) -221.98100 (maps) -221.98400 (te) 20 (xtual) -221.98300 (queries) -221.98300 (and) -223.00700 (visual) -222.01400 (featur) 37.00120 (es) ] TJ 11.95470 TL T* [ (fr) 44.98640 (om) -398.98400 (various) -399.00100 (r) 37.01960 (e) 39.98840 (gions) -399.00300 (into) -397.98800 (a) -398.98700 (shar) 36.99630 (ed) -398.99100 (space) -398.98500 (wher) 36.99380 (e) -398.99700 (the) 30.00230 (y) -398.99700 (ar) 36.98650 (e) ] TJ T* [ (compar) 36.99380 (ed) -247.99700 (for) -248.01900 (r) 37.01960 (ele) 15.01590 (vance) -248 (with) -248.00400 (an) -247.98800 (inner) -247 (pr) 44.98510 (oduct\056) -310.00700 (Our) -247.98300 (method) ] TJ T* [ (e) 19.99240 (xhibits) -436.01200 <7369676e690263616e74> -435.01700 (impr) 44.99370 (o) 10.00320 (vements) -435.99700 (in) -435.01300 (answering) -435.99600 (questions) ] TJ T* [ (suc) 14.98460 (h) -451.99800 (as) -451.01200 (\223what) -451.99700 (color) 110.98000 (\054) 140.01800 (\224) -502.01900 (wher) 36.99380 (e) -452.00800 (it) -451.98800 (is) -451.00900 (necessary) -451.99300 (to) -451.99100 (e) 15.01220 (valuate) ] TJ T* [ (a) -346.99600 <73706563690263> -346.99100 (location\054) -372.01600 (and) -346.98600 (\223what) -346.99500 (r) 45.01820 (oom\054) 139.99300 (\224) -370.98200 (wher) 36.99380 (e) -347.98600 (it) -346.98600 (selectively) ] TJ 11.95630 TL T* [ <6964656e7469026573> -363.98400 (informative) -364.99300 (ima) 10.01300 (g) 10.00320 (e) -363.98300 (r) 37.01960 (e) 39.98840 (gions\056) -652.99800 (Our) -364.98300 (model) -364.00300 (is) -364.98600 (tested) ] TJ 11.95510 TL T* [ (on) -376.00600 (the) -374.98800 (r) 37.01960 (ecently) -376.01300 (r) 37.01960 (eleased) -376.00600 (VQA) -374.98500 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R20 9.96260 Tf 1 0 0 1 179.30700 401.33600 Tm (1) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R20 9.96260 Tf 1 0 0 1 184.28800 401.33600 Tm [ (\135) -376.01000 (dataset\054) -406.99600 (whic) 15 (h) -376.01100 (featur) 37.00120 (es) ] TJ -134.17600 -11.95470 Td [ (fr) 36.98770 (ee\055form) -249.98800 (human\055annotated) -250.01600 (questions) -249.99500 (and) -249.98300 (answer) 9.99098 (s\056) ] TJ /R12 11.95520 Tf 37.38200 TL T* [ (1\056) -249.99000 (Intr) 18.01460 (oduction) ] TJ /R14 9.96260 Tf 11.95510 -19.43520 Td [ (V) 59.99310 (isual) -268.99300 (question) -269.01800 (answering) -270.01400 (\050VQA\051) -269.01100 (is) -269.00200 (the) -269.00700 (task) -269 (of) -269.99200 (answer) 20 (\055) ] TJ -11.95510 -11.95510 Td [ (ing) -463.00400 (a) -462.98600 (natural) -463.00900 (language) -463.00600 (question) -462.98500 (about) -463.00400 (an) -462.98200 (image\056) -949 (VQA) ] TJ 11.95470 TL T* [ (includes) -409.00300 (man) 14.99010 (y) -410.00500 (challenges) -409.01100 (in) -409.01800 (language) -409.01500 (representation) -409.98600 (and) ] TJ T* [ (grounding\054) -275.01000 (recognition\054) -275.01500 (comm) 1.00473 (on) -269.98400 (sense) -269.98700 (reasoning\054) -274.98800 (and) -269.98900 (spe\055) ] TJ 11.95590 TL T* [ (cialized) -314.99700 (tasks) -316.00600 (lik) 10.01790 (e) -315.01100 (counting) -315.99000 (and) -315 (reading\056) -505.98200 (In) -315.98400 (this) -314.98200 (paper) 39.99090 (\054) -331.98100 (we) ] TJ 11.95510 TL T* [ (focus) -223.99300 (on) -225.01100 (a) -224.00600 (k) 10.00320 (e) 15.01280 (y) -225.01600 (probl) 0.99861 (em) -225.01400 (for) -224.00300 (VQA) -225.01300 (and) -223.99600 (other) -223.99200 (visual) -224.99400 (reasoning) ] TJ T* [ (tasks\072) -345.99400 (kno) 24.99150 (wing) -268.99400 (where) -267.99800 (to) -267.98200 (look\056) -365.01300 (Consider) -268 (Figure) ] TJ ET Q 1 0 0 rg q 10 0 0 10 0 0 cm BT /R14 9.96260 Tf 1 0 0 1 242.15100 260.83300 Tm (1) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R14 9.96260 Tf 1 0 0 1 247.13200 260.83300 Tm [ (\056) -364.99500 (It\047) 55.01040 (s) -267.99000 (easy) ] TJ -197.02000 -11.95510 Td [ (to) -227.00900 (answer) -227.01900 (\223What) -227.99700 (color) -227.01100 (is) -227.00700 (the) -227.01400 (w) 10 (alk) -227.01400 (light\077\224) -302.98600 (if) -227.01600 (the) -227.01400 (light) -226.98900 (b) 20.00160 (ulb) -227.98400 (is) ] TJ 11.95510 TL T* [ (localized\054) -213.00700 (whi) 0.99493 (le) -204.00300 (answering) -203.00500 (whether) -203.01600 (it\047) 54.98220 (s) -203.01900 (raining) -203.00300 (may) -203.99300 (be) -202.98600 (dealt) ] TJ T* [ (with) -224.01900 (by) -223.99200 (identifying) -223.99400 (umbrellas\054) -228.98800 (puddles\054) -228.98900 (or) -224 (cloudy) -223 (skies\056) -302.00600 (W) 79.98660 (e) ] TJ 11.95630 TL T* [ (w) 10.00140 (ant) -259.98800 (to) -259.98400 (learn) -259.99600 (where) -258.98000 (to) -259.98400 (look) -260.01300 (to) -259.98400 (answer) -259.99300 (questions) -259.99300 (supervised) ] TJ 11.95510 TL T* [ (by) -279.00200 (only) -278.99000 (images) -278.99000 (and) -279.00600 (question\057answer) -279.00600 (pairs\056) -397.02000 (F) 14.99260 (or) -279.01000 (e) 15.01220 (xample\054) -286.98900 (if) ] TJ T* [ (we) -253.01800 (ha) 19.99730 (v) 14.98280 (e) -252.98200 (se) 25.01730 (v) 14.98280 (eral) -254.00200 (traini) 0.99861 (ng) -253.98700 (e) 15.01220 (xamples) -252.99000 (for) -253.01800 (\223What) -253.01200 (time) -252.99500 (of) -253.99400 (day) -253.01200 (is) ] TJ T* [ (it\077\224) -307.00200 (or) -241.99700 (similar) -242.00600 (questions\054) -243.01200 (the) -241.99100 (system) -240.98700 (should) -242.00400 (learn) -241.99900 (what) -241.98900 (kind) ] TJ T* [ (of) -228 (answer) -229.01900 (is) -227.98800 (e) 15.01220 (xpected) -227.99300 (and) -229.01400 (where) -228.00500 (in) -227.98900 (the) -229.01300 (image) -227.98600 (it) -227.98600 (should) -228.98600 (base) ] TJ 11.95590 TL T* [ (its) -250.01600 (response\056) ] TJ 11.95510 -12.46020 Td [ (Learning) -360.00400 (where) -359.00300 (to) -360.00600 (look) -359.99600 (from) -358.98800 (question\055image) -360.00900 (pairs) -360.00900 (has) ] TJ -11.95510 -11.95510 Td [ (man) 14.99080 (y) -319.98100 (challenges\056) -520.01000 (Questions) -319.98500 (such) -320.01300 (as) -319.98500 (\223What) -319.98100 (sport) -320.00500 (is) -320.01300 (this\077\224) ] TJ T* [ (might) -325.00700 (be) -325.00400 (best) -324.99100 (answered) -325.00200 (usi) 0.99861 (ng) -324.99500 (the) -324.99700 (full) -324.98900 (image\056) -534.99200 (Other) -324.99700 (ques\055) ] TJ 11.95510 TL T* [ (tions) -410.01900 (such) -411.01800 (as) -410.00900 (\223What) -410.00600 (is) -411.01600 (on) -410 (the) -410.98300 (sof) 10.00570 (a\077\224) -791.01800 (or) -410.00800 (\223What) -410.00500 (color) -410.98000 (is) ] TJ T* [ (the) -263.98700 (w) 10.00140 (oman\047) 54.98590 (s) -263.98900 (shirt\077\224) -352.98600 (require) -263.99300 (focusing) -264.00900 (on) -265.00500 (particul) 1 (ar) -264.98300 (re) 15.00980 (gions\056) ] TJ T* [ (Still) -299.98500 (others) -300.99300 (such) -300.01600 (as) -299.98800 (\223What) -301.00500 (does) -300.01600 (the) -299.98200 (sign) -300.98700 (say\077\224) -460.98400 (or) -299.98700 (\223) 79.98180 (Are) -301.01300 (the) ] TJ ET Q q 3088.62000 3396.17000 2362.46000 2036.95000 re W n /R42 Do Q /R36 gs q 10 0 0 10 0 0 cm BT /R14 8.96640 Tf 1 0 0 1 308.86200 328.65800 Tm [ (Figure) -328.00900 (1\056) -327.98200 (Our) -327.98500 (goal) -329.01700 (i) 1.01017 (s) -329 (to) -327.99600 (identify) -328.00900 (the) -328.01500 (correct) -327.99300 (answer) -328.02000 (for) -327.98800 (a) -327.97700 (natural) ] TJ 10.95780 TL T* [ (language) -215 (question\054) -221.98500 (such) -215.01200 (as) -214.00700 (\223What) -215 (color) -215.01700 (is) -214.98400 (the) -215.00600 (w) 10.00920 (alk) -215.00600 (light) 0.98567 (\077\224) -298.98400 (or) -215.01200 (\223Is) -213.97700 (it) ] TJ 10.95900 TL T* [ (raining\077\224) -344.00300 (W) 80.01060 (e) -262.02200 (particularly) -261.00600 (focus) -260.98400 (on) -262.00200 (the) -261.01100 (problem) -262.01900 (of) -261.01700 (learning) -260.99000 (where) ] TJ T* [ (to) -208.01700 (look\056) -295.99700 (This) -209.01900 (is) -208.01100 (a) -208 (challenging) -208.01400 (problem) -207.99800 (as) -209 (it) -208.00900 (requires) -207.99200 (grounding) -207.98700 (lan\055) ] TJ T* [ (guage) -249.01400 (with) -249.01200 (vision) -248.99800 (and) -248.99500 (learning) -249.00900 (to) -249.00900 (recognize) -249.01400 (objects\054) -249.01400 (use) -248.99300 (relations\054) ] TJ T* [ (and) -301.01300 (determ) 1.01017 (ine) -301.00400 (rele) 25.00660 (v) 25.00660 (ance\056) -462.00300 (F) 15.00560 (or) -300.00800 (e) 15.01370 (xample\054) -313.01800 (whether) -301.00400 (it) -300.01900 (is) -300.98300 (raining) -300.00800 (may) ] TJ T* [ (be) -245.98900 (determined) -247.01600 (by) -246.01400 (detecting) -247.00200 (the) -245.98100 (presence) -246.01400 (of) -246.98900 (puddles\054) -247 (gray) -246.00300 (skies\054) -246.99400 (or) ] TJ T* [ (umbrellas) -212.01100 (in) -210.97900 (the) -212 (scene\054) -218.99000 (whereas) -211.97800 (the) -212 (color) -212.01100 (of) -211.00400 (the) -212 (w) 10.00920 (alk) -212 (light) -211.00400 (requires) ] TJ T* [ (focused) -310.01700 (attention) -311.01600 (on) -310.01200 (the) -310.02200 (light) -309.98400 (alone\056) -490.98800 (The) -310.99200 (abo) 15.00830 (v) 14.98650 (e) -309.98400 <0267757265> -309.98400 (sho) 25.01200 (ws) -310.01700 (e) 15.01370 (x\055) ] TJ T* [ (ample) -249.99700 (attention) -249.98100 (re) 14.99740 (gions) -250.00800 (produced) -249.98900 (by) -249.97800 (our) -249.99500 (proposed) -250.01100 (model\056) ] TJ /R14 9.96260 Tf 29.47620 TL T* [ (man) -378.00300 (and) -378.99100 (w) 10 (oman) -377.99800 (dating\077\224) -695.98700 (require) -378.01500 (specialized) -379.02000 (kno) 24.99090 (wledge) ] TJ 11.95510 TL T* [ (or) -293.00700 (reasoning) -292.98800 (that) -294.01500 (we) -293.01200 (do) -293 (not) -292.99000 (e) 15.01220 (xpect) -293.98000 (to) -293 (achie) 25.01540 (v) 14.98280 (e\056) -438.98600 (The) -293.98500 (system) ] TJ 11.95590 TL T* [ (needs) -336.99500 (to) -336.99000 (learn) -337.00300 (to) -338.01000 (re) 1.01209 (cognize) -338.01200 (objects\054) -359.01400 (infer) -337.00300 (spatial) -337.01700 (relations\054) ] TJ 11.95510 TL T* [ (determine) -252.99700 (rele) 25.01050 (v) 24.98110 (ance\054) -254.98200 (and) -253.01200 <026e64> -254.01200 (correspondence) -253.00700 (between) -253.99200 (nat\055) ] TJ T* [ (ural) -245.99300 (language) -247.00300 (and) -245.99300 (visual) -246.99100 (features\056) -307.98800 (Our) -247.01300 (k) 10.00320 (e) 15.01220 (y) -245.99300 (idea) -247.01300 (is) -245.98300 (to) -245.98800 (learn) -247.01800 (a) ] TJ T* [ (non\055linear) -302.98900 (mapping) -303.01300 (of) -303.98600 (language) -302.99400 (and) -303.00300 (visual) -302.98400 (re) 15.00730 (gion) -304.00300 (features) ] TJ T* [ (into) -385.00900 (a) -385 (common) -384.01900 (latent) -385.01700 (space) -384.99500 (to) -384.98500 (determine) -383.99500 (rele) 25.01050 (v) 24.98110 (ance\056) -714.98400 (The) ] TJ T* [ (rele) 25.01050 (v) 24.98110 (ant) -303 (re) 15.00980 (gions) -303.01800 (are) -304 (then) -302.99600 (used) -302.99800 (to) -302.99800 (score) -303.00800 (a) -303.99300 <73706563690263> -302.98900 (question\055) ] TJ 11.95590 TL T* [ (answer) -222 (pairing\056) -300.98400 (The) -223.01700 (latent) -221.98300 (embedding) -222.00700 (and) -223.01700 (the) -221.99200 (scoring) -223.00200 (func\055) ] TJ 11.95510 TL T* [ (tion) -413.98500 (are) -413.98200 (learned) -414.00900 (joint) 1.01209 (ly) -413.99700 (using) -413.98000 (a) -414.01400 (mar) 18.00190 (gin\055based) -413.99000 (loss) -413.98500 (super) 20.00650 (\055) ] TJ T* [ (vised) -314.98700 (solely) -315.01900 (by) -316.01600 (question\055answer) -315.01100 (pairings\056) -504.98400 (W) 79.98660 (e) -315.01100 (perform) -316.01600 (e) 15.01220 (x\055) ] 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 32 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 35 Tm (4613) 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
/Interpolate true
/Height 426
/Length 12623
/ColorSpace /DeviceRGB
/Width 640
/BitsPerComponent 8
/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
$4%&'()*56789:CDEFGHIJSTUVWXYZcdefghijstuvwxyz ? (
(Q(((ES 4 QE RR(( ( (@QE QE0S4S 8TAHxi:PqRFHx~4$_rrՔˍň5Z