%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 12 0 R ] /Type /Pages /Count 10 >> endobj 2 0 obj << /Title (Improved Fusion of Visual and Language Representations by Dense Symmetric Co\055Attention for Visual Question Answering) /Producer (PyPDF2) /Author (Duy\055Kien Nguyen\054 Takayuki Okatani) /Subject (2018 IEEE Conference on Computer Vision and Pattern Recognition) >> endobj 3 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 14 0 R /Resources << /XObject << /x8 15 0 R /x6 18 0 R /x12 21 0 R /x10 24 0 R >> /ExtGState << /s9 27 0 R /s11 30 0 R /R16 33 0 R /a0 << /CA 1 /ca 1 >> /s5 34 0 R /s7 37 0 R >> /Font << /R29 40 0 R /F2 44 0 R /F1 45 0 R /R17 46 0 R /R21 49 0 R /R19 52 0 R /R23 56 0 R /R25 60 0 R /R27 64 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 67 0 R 68 0 R 69 0 R 70 0 R 71 0 R 72 0 R 73 0 R 74 0 R ] >> endobj 4 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 75 0 R /Resources << /ColorSpace << /R63 76 0 R >> /ExtGState << /R64 78 0 R /R61 79 0 R /R16 33 0 R >> /Font << /R65 80 0 R /R29 40 0 R /R67 84 0 R /F1 88 0 R /R17 46 0 R /R19 52 0 R /R71 89 0 R /R69 92 0 R /F2 96 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 97 0 R 98 0 R 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 ] >> endobj 5 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 122 0 R /Resources << /ColorSpace << /R63 76 0 R >> /ExtGState << /R64 78 0 R /R61 79 0 R /R16 33 0 R >> /Font << /R108 123 0 R /R71 89 0 R /R110 126 0 R /R102 130 0 R /R104 134 0 R /F2 138 0 R /R106 139 0 R /R65 80 0 R /F1 144 0 R /R29 40 0 R /R17 46 0 R /R19 52 0 R /R69 92 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 145 0 R 146 0 R 147 0 R 148 0 R 149 0 R 150 0 R 151 0 R 152 0 R 153 0 R 154 0 R 155 0 R 156 0 R 157 0 R 158 0 R 159 0 R 160 0 R 161 0 R 162 0 R 163 0 R 164 0 R ] >> endobj 6 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 165 0 R /Resources << /ColorSpace << /R63 76 0 R >> /ExtGState << /R127 166 0 R /R126 167 0 R /R64 78 0 R >> /Font << /R71 89 0 R /R65 80 0 R /R130 168 0 R /R132 172 0 R /F1 175 0 R /R134 176 0 R /R136 179 0 R /R138 183 0 R /F2 186 0 R /R128 187 0 R /R108 123 0 R /R140 191 0 R /R102 130 0 R /R104 134 0 R /R106 139 0 R /R29 40 0 R /R17 46 0 R /R21 49 0 R /R19 52 0 R /R23 56 0 R /R69 92 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 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 ] >> endobj 7 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 203 0 R /Resources << /ExtGState << /R16 33 0 R >> /Font << /R108 123 0 R /R71 89 0 R /R102 130 0 R /R104 134 0 R /F2 204 0 R /R106 139 0 R /R65 80 0 R /R132 172 0 R /F1 205 0 R /R134 176 0 R /R29 40 0 R /R136 179 0 R /R17 46 0 R /R19 52 0 R /R69 92 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 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 << /ExtGState << /R16 33 0 R >> /Font << /F2 212 0 R /R102 130 0 R /R104 134 0 R /R29 40 0 R /F1 213 0 R /R17 46 0 R /R19 52 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 214 0 R 215 0 R 216 0 R 217 0 R 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 ] >> endobj 9 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 245 0 R /Resources << /XObject << /R202 246 0 R /R203 247 0 R /R200 248 0 R /R201 249 0 R /R204 250 0 R /R198 251 0 R /R199 252 0 R >> /ExtGState << /R16 33 0 R >> /Font << /F2 253 0 R /R71 89 0 R /R102 130 0 R /R104 134 0 R /R65 80 0 R /R132 172 0 R /F1 254 0 R /R17 46 0 R /R19 52 0 R /R69 92 0 R >> /ProcSet [ /Text /ImageC /ImageB /PDF /ImageI ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 255 0 R 256 0 R 257 0 R 258 0 R ] >> endobj 10 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 259 0 R /Resources << /ExtGState << /R16 33 0 R >> /Font << /R65 80 0 R /R71 89 0 R /R102 130 0 R /R104 134 0 R /F2 260 0 R /R29 40 0 R /F1 261 0 R /R17 46 0 R /R19 52 0 R /R69 92 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 262 0 R 263 0 R 264 0 R 265 0 R 266 0 R 267 0 R 268 0 R 269 0 R 270 0 R 271 0 R 272 0 R 273 0 R 274 0 R 275 0 R 276 0 R 277 0 R 278 0 R 279 0 R 280 0 R 281 0 R 282 0 R 283 0 R 284 0 R ] >> endobj 11 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 285 0 R /Resources << /ExtGState << /R16 33 0 R >> /Font << /R29 40 0 R /R19 52 0 R /F1 286 0 R /F2 287 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 288 0 R 289 0 R 290 0 R 291 0 R 292 0 R 293 0 R 294 0 R 295 0 R 296 0 R 297 0 R 298 0 R 299 0 R 300 0 R 301 0 R 302 0 R 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 336 0 R 337 0 R 338 0 R 339 0 R 340 0 R 341 0 R 342 0 R 343 0 R 344 0 R 345 0 R 346 0 R 347 0 R 348 0 R 349 0 R 350 0 R ] >> endobj 12 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 351 0 R /Resources << /ExtGState << /R16 33 0 R >> /Font << /R29 40 0 R /R19 52 0 R /F1 352 0 R /F2 353 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 354 0 R 355 0 R 356 0 R 357 0 R 358 0 R ] >> endobj 13 0 obj << /Type /Catalog /Pages 1 0 R >> endobj 14 0 obj << /Length 17083 >> stream q q q 0.1 0 0 0.1 0 0 cm /R16 gs 0 g q 10 0 0 10 0 0 cm BT /R17 14.3462 Tf 1 0 0 1 119.735 675.175 Tm [ (Impr) 17.9921 (o) 9.99625 (v) 9.99625 (ed) -250.004 (Fusion) -250.013 (of) -250.012 (V) 37.0137 (isual) -250.005 (and) -249.991 (Language) -250.01 (Repr) 18.0014 (esentations) ] TJ -23.1043 -17.825 Td [ (by) -250.013 (Dense) -249.995 (Symmetric) -250.01 (Co\055Attention) -250.002 (f) 24.9923 (or) -249.995 (V) 37.0137 (isual) -250.005 (Question) -249.998 (Answering) ] TJ /R19 11.9552 Tf 90.0102 -47.7371 Td [ (Duy\055Kien) -250.002 (Nguyen) ] TJ /R21 7.9701 Tf 88.6469 4.33828 Td [ (1) -0.30019 ] TJ /R19 11.9552 Tf 10.7102 -4.33828 Td [ (and) -749.998 (T) 80.0147 (akayuki) -250.008 (Okatani) ] TJ /R21 7.9701 Tf 111.266 4.33828 Td [ (1) -0.30019 ] TJ /R23 7.9701 Tf 4.23398 0 Td [ (\073) -0.10109 ] TJ /R21 7.9701 Tf 2.35312 0 Td [ (2) -0.30019 ] TJ -219.828 -15.323 Td [ (1) -0.29866 ] TJ /R19 11.9552 Tf 4.73203 -4.33828 Td [ (T) 80.0137 (ohoku) -250.014 (Uni) 24.9957 (v) 14.9851 (ersity) ] TJ /R21 7.9701 Tf 107.154 4.33828 Td [ (2) -0.30019 ] TJ /R19 11.9552 Tf 4.73203 -4.33828 Td [ (RIKEN) -250.014 (Center) -249.997 (for) -249.995 (AIP) ] TJ /R25 8.9664 Tf -104.474 -13.8637 Td [ (f) -0.8999 ] TJ /R27 8.9664 Tf 4.6082 0 Td [ (kien\054) -599.986 (okatani) ] TJ /R25 8.9664 Tf 69.9379 0 Td [ (g) -0.90126 ] TJ /R27 8.9664 Tf 4.6082 0 Td (\100vision\056is\056tohoku\056ac\056jp) Tj /R17 11.9552 Tf -129.326 -49.066 Td (Abstract) Tj /R29 9.9626 Tf -83.9281 -24.4211 Td [ (A) -231.997 (k) 9.99404 (e) 29.9907 (y) -230.985 (solution) -231.992 (to) -231.008 (visual) -232.014 (question) -232.004 (answering) -231.01 (\050VQA\051) -231.996 (e) 19.9918 (xists) ] TJ -11.9547 -11.884 Td [ (in) -286.018 (how) -286.007 (to) -286.018 (fuse) -286.017 (visual) -286.005 (and) -286.016 (langua) 9.99098 (g) 10.0032 (e) -285.996 (featur) 37.0012 (es) -285.989 (e) 19.9918 (xtr) 14.9877 (acted) -285.984 (fr) 44.9851 (om) ] TJ 11.8832 TL T* [ (an) -380.005 (input) -379.986 (ima) 10.013 (g) 10.0032 (e) -380.02 (and) -381.02 (question\056) -700 (W) 91.9859 (e) -380.02 (show) -379.986 (that) -381.01 (an) -380.005 (attention) ] TJ 11.884 TL T* [ (mec) 15.011 (hanism) -369.005 (that) -369.012 (enables) -369.997 (dense) 10.0057 (\054) -398.99 (bi\055dir) 37.0012 (ectional) -369.009 (inter) 15 (actions) ] TJ 11.8828 TL T* [ (between) -360.981 (the) -360.991 (two) -361.012 (modalities) -360.99 (contrib) 19.9881 (utes) -360.984 (to) -362.006 (boost) -361.008 (accur) 14.9852 (acy) ] TJ 11.884 TL T* [ (of) -392.981 (pr) 36.9852 (e) 1.01515 (d) -1.01454 (i) 0.98758 (ction) -393.012 (of) -392.981 (answer) 11.0092 (s\056) -737.986 <53706563690263616c6c79> 54.9957 (\054) -428.005 (we) -392.983 (pr) 36.9865 (esent) -392.003 (a) -392.988 (sim\055) ] TJ 11.8832 TL T* [ (ple) -301.981 (ar) 36.9852 (c) 15.0128 (hitectur) 37.0036 (e) -301.994 (that) -302.004 (is) -302.015 (fully) -302.006 (symmetric) -302.008 (between) -302.011 (visual) -302.001 (and) ] TJ 11.884 TL T* [ (langua) 9.99159 (g) 10.0032 (e) -293.994 (r) 37.0196 (epr) 36.9816 (esentations\054) -305 (in) -294.017 (whic) 14.9987 (h) -293.985 (eac) 15.0147 (h) -295.005 (ques) 0.99248 (tion) -294.985 (wor) 36.9987 (d) -293.985 (at\055) ] TJ 11.8828 TL T* [ (tends) -336.984 (on) -338.012 (ima) 10.013 (g) 10.0032 (e) -337.007 (r) 37.0196 (e) 39.9884 (gions) -337.014 (and) -336.988 (eac) 15.0134 (h) -338.017 (ima) 10.013 (g) 10.0032 (e) -337.007 (r) 37.0183 (e) 39.9884 (gion) -337.02 (attends) -337.99 (on) ] TJ 11.884 TL T* [ (question) -331.987 (wor) 36.9987 (ds\056) -556.011 (It) -332.019 (can) -332.017 (be) -331.983 (stac) 20.0101 (k) 9.99343 (ed) -331.983 (to) -332.011 (form) -332 (a) -332.018 (hier) 14.9926 (ar) 36.9865 (c) 15.0122 (hy) -331.984 (for) ] TJ 11.8832 TL T* [ (multi\055step) -245.998 (inter) 14.9987 (actions) -244.987 (between) -245.981 (an) -245.988 (ima) 10.013 (g) 10.0032 (e\055quest) 0.99248 (ion) -245.981 (pair) 110.984 (\056) -309.005 (W) 91.9871 (e) ] TJ 11.884 TL T* [ (show) -412.019 (thr) 44.9913 (ough) -411.01 (e) 19.9918 (xperiments) -411.983 (that) -411.005 (the) -412.002 (pr) 44.9851 (oposed) -411.988 (ar) 36.9865 (c) 15.0122 (hitectur) 37.0036 (e) ] TJ 11.8828 TL T* [ (ac) 15.0177 (hie) 14.9859 (ves) -359.001 (a) -360.013 (ne) 15.0183 (w) -358.985 (state\055of\055the\055art) -360.007 (on) -358.989 (VQA) -358.989 (and) -360.004 (VQA) -358.989 (2\0560) -360.006 (de\055) ] TJ T* [ (spite) -292.009 (its) -292.99 (small) -292.002 (size) 14.9926 (\056) -436.983 (W) 91.9859 (e) -293.015 (also) -292.006 (pr) 36.9852 (esent) -291.981 (qualitative) -293.003 (e) 15.0122 (valuation\054) ] TJ 11.884 TL T* [ (demonstr) 15.011 (ating) -317.016 (how) -316.983 (the) -316.018 (pr) 44.9839 (oposed) -316.984 (attention) -317.001 (mec) 15.0122 (hanism) -317.013 (can) ] TJ 11.8828 TL T* [ (g) 10.0032 (ener) 15.0196 (ate) -384.006 (r) 37.0196 (easonable) -383.985 (attention) -382.99 (maps) -383.996 (on) -384.005 (ima) 10.013 (g) 10.0032 (es) -384.012 (and) -384 (ques\055) ] TJ 11.884 TL T* [ (tions\054) -250.004 (whic) 14.9987 (h) -249.993 (leads) -250.018 (to) -249.985 (the) -249.99 (corr) 36.9865 (ect) -249.999 (answer) -250.005 (pr) 36.9865 (ediction\056) ] TJ /R17 11.9552 Tf 37.5289 TL T* [ (1\056) -249.99 (Intr) 18.0146 (oduction) ] TJ /R19 9.9626 Tf 11.9547 -19.4391 Td [ (There) -285.014 (has) -285.005 (been) -285.01 (a) -285.016 <7369676e690263616e74> -285.003 (progress) -284.994 (in) -285 (the) -285.984 (study) -284.982 (of) -285.009 (vi\055) ] TJ -11.9547 -11.884 Td [ (sual) -267.98 (question) -269.018 (answering) -268.014 (\050VQA\051) -267.992 (o) 14.9828 (v) 14.9828 (er) -268.002 (a) -269.019 (short) -268.014 (period) -268.019 (of) -269.011 (time) ] TJ 11.8832 TL T* [ (since) -299 (its) -298.008 (introduction\054) -310.992 (sho) 24.9922 (wing) -298.989 (rapid) -298.019 (boost) -298.98 (of) -299.006 (performance) ] TJ 11.884 TL T* [ (for) -359.98 (common) -359.983 (benchmark) -360.983 (datasets\056) -639.99 (This) -359.996 (progress) -360.001 (has) -360.991 (been) ] TJ 11.8828 TL T* [ (mainly) -345.008 (brought) -345.011 (about) -344.983 (by) -344.991 (tw) 10.0081 (o) -344.996 (lines) -345.02 (of) -345 (research\054) -369.007 (the) -344.994 (de) 25.0154 (v) 14.9828 (el\055) ] TJ 11.884 TL T* [ (opment) -408.02 (of) -408.009 (better) -408.983 (attention) -408.014 (mechanisms) -408.008 (and) -408.005 (the) -408.983 (impro) 15.0048 (v) 14.9828 (e\055) ] TJ 11.8832 TL T* [ (ment) -377.995 (in) -376.984 (fusion) -377.991 (of) -378.014 (features) -378.017 (e) 15.0122 (xtracted) -376.987 (from) -378.006 (an) -378.015 (input) -377.986 (image) ] TJ 11.884 TL T* [ (and) -249.992 (question\056) ] TJ 11.9547 -12.466 Td [ (Since) -348.983 (introduced) -349.013 (by) -348.011 (Bahdanau) -349.019 (et) -348.998 (al\056) -606.012 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R19 9.9626 Tf 1 0 0 1 220.223 164.184 Tm (3) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R19 9.9626 Tf 1 0 0 1 225.205 164.184 Tm [ (\135\054) -374.016 (attention) -349.005 (has) ] TJ -175.093 -11.8828 Td [ (been) -237.018 (playing) -236.009 (an) -237.018 (important) -237.017 (role) -235.995 (in) -237.007 (solutions) -235.985 (of) -237.017 (v) 24.9811 (arious) -237.005 (prob\055) ] TJ 11.884 TL T* [ (lems) -290.989 (of) -291.008 <61727469026369616c> -290.983 (i) 0.98758 (ntelligence) -290.983 (ranging) -290.991 (from) -291 (tasks) -290.991 (using) -290.981 (sin\055) ] TJ 11.8832 TL T* [ (gle) -386.006 (modality) -386.009 (\050e\056g\056\054) -420.003 (l) 0.98758 (anguage\054) -419.999 (speech\054) -420.002 (and) -386.009 (vision\051) -385.982 (to) -386.002 (mul\055) ] TJ 11.884 TL T* [ (timodal) -341.002 (tasks\056) -581.998 (F) 14.9926 (or) -341 (VQA\054) -340.991 (attention) -339.986 (on) -340.992 (image) -340.987 (re) 15.0098 (gions) -341.012 (gen\055) ] TJ 11.8828 TL T* [ (erated) -452.007 (from) -451.993 (the) -451.996 (input) -452.013 (question) -452.007 (w) 10 (as) -452.002 <02727374> -452.018 (introduced) -452.015 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R19 9.9626 Tf 1 0 0 1 273.086 104.767 Tm (28) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R19 9.9626 Tf 1 0 0 1 283.048 104.767 Tm (\135) Tj -232.936 -11.884 Td [ (and) -247.012 (then) -247.985 (se) 25.0173 (v) 14.9828 (eral) -246.983 (e) 15.0122 (xtensions) -247.015 (ha) 19.9979 (v) 14.9828 (e) -246.983 (been) -247.997 (proposed) -246.993 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R19 9.9626 Tf 1 0 0 1 245.738 92.8832 Tm (17) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R19 9.9626 Tf 1 0 0 1 255.7 92.8832 Tm (\054) Tj ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R19 9.9626 Tf 1 0 0 1 260.652 92.8832 Tm (31) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R19 9.9626 Tf 1 0 0 1 270.614 92.8832 Tm (\054) Tj ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R19 9.9626 Tf 1 0 0 1 275.575 92.8832 Tm (4) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R19 9.9626 Tf 1 0 0 1 280.557 92.8832 Tm (\135\056) Tj -230.445 -11.8832 Td [ (Meanwhile\054) -270.006 (researchers) -266.99 (ha) 19.9979 (v) 14.9828 (e) -266 (proposed) -266.01 (se) 25.0179 (v) 14.9828 (eral) -266 (methods) -267.017 (for) ] TJ 258.75 450.36 Td [ (feature) -347.988 (fusion) -349.015 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R19 9.9626 Tf 1 0 0 1 371.686 531.36 Tm (5) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R19 9.9626 Tf 1 0 0 1 376.668 531.36 Tm (\054) Tj ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R19 9.9626 Tf 1 0 0 1 382.625 531.36 Tm (14) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R19 9.9626 Tf 1 0 0 1 392.588 531.36 Tm (\054) Tj ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R19 9.9626 Tf 1 0 0 1 398.546 531.36 Tm (32) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R19 9.9626 Tf 1 0 0 1 408.508 531.36 Tm [ (\135\054) -372.997 (where) -349.005 (the) -348.011 (aim) -348.986 (is) -348.006 (to) -348.011 (obtain) -349.015 (bet\055) ] TJ -99.6461 -11.884 Td [ (ter) -257.996 (fused) -257.009 (representation) -258.011 (of) -257.016 (image) -257.981 (and) -257.011 (question) -258.001 (pairs\056) -332.989 (These) ] TJ 11.8832 TL T* [ (studies) -248.985 (updated) -250.017 (the) -249.01 (state\055of\055the\055art) -249.988 (for) -249.017 (common) -250.002 (benchmark) ] TJ 11.884 TL T* [ (datasets) -250.017 (at) -249.995 (the) -249.99 (time) -250.015 (of) -249.995 (each) -250.007 (publication\056) ] TJ 11.9547 -15.4398 Td [ (W) 79.9866 (e) -216.008 (observ) 14.9926 (e) -216.008 (that) -215.988 (these) -215.993 (tw) 10.0081 (o) -215.998 (lines) -215.984 (of) -216.003 (research) -216.998 (ha) 19.9967 (v) 14.9828 (e) -216.008 (been) -216.003 (in\055) ] TJ -11.9547 -11.884 Td [ (dependently) -297.017 (conducted) -298.009 (so) -297 (f) 9.99343 (ar) 55 (\056) -451.998 (This) -298.004 (is) -296.995 (particularly) -296.985 (the) -297.98 (case) ] TJ 11.8832 TL T* [ (with) -280.01 (the) -279.005 (studies) -280 (of) -279.99 (feature) -279.017 (fusion) -280.007 (methods\054) -287.011 (where) -279.997 (attention) ] TJ 11.884 TL T* [ (is) -358.006 (considered) -358.002 (to) -358.006 (be) -358.019 (optional\054) -384.995 (e) 25.0105 (v) 14.9828 (en) -358.019 (though) -359.004 (the) -358.009 (best) -358.004 (perfor) 19.9918 (\055) ] TJ 11.8828 TL T* [ (mance) -275.988 (is) -276.02 (achie) 25.0154 (v) 14.9828 (ed) -275.993 (with) -276.01 (it\056) -387.999 (Ho) 24.986 (we) 25.0154 (v) 14.9828 (er) 39.9835 (\054) -283.007 (we) -275.993 (think) -276.003 (that) -276.018 (the) 14.9877 (y) -275.988 (are) ] TJ 11.884 TL (rather) ' /R29 9.9626 Tf 26.334 0 Td [ (two) -310.982 (dif) 18.0166 (fer) 36.984 (ent) -311 (appr) 44.9949 (oac) 14.9828 (hes) -312.021 (towar) 37.0085 (ds) -310.997 (the) -310.997 (same) -312.017 (goal) ] TJ /R19 9.9626 Tf 194.206 0 Td [ (\056) -493.991 (In) ] TJ -220.54 -11.8832 Td [ (particular) 40.0056 (\054) -247.015 (we) -246 (ar) 17.9896 (gue) -247.013 (that) -245.983 (a) -246.983 (better) -245.988 (attention) -246.003 (mechanism) -246.983 (leads) ] TJ T* [ (to) -249.985 (a) -250.002 (better) -249.99 (fused) -249.99 (representation) -250.012 (of) -249.997 (image\055question) -249.988 (pairs\056) ] TJ 11.9547 -15.4398 Td [ (Moti) 25.0056 (v) 24.9811 (ated) -561.982 (by) -562.014 (this\054) -639.984 (we) -561.985 (propose) -562.984 (a) -561.99 (no) 14.9877 (v) 14.9828 (el) -561.98 (co\055attention) ] TJ -11.9547 -11.884 Td [ (mechanism) -224.987 (for) -226.004 (impro) 15.0048 (v) 14.9828 (ed) -224.982 (fusion) -226.016 (of) -225.021 (visual) -226.016 (and) -225.016 (language) -225.987 (rep\055) ] TJ 11.8832 TL T* [ (resentations\056) -305.013 (Gi) 24.986 (v) 14.9828 (en) -236 (representations) -234.99 (of) -236 (an) -235.02 (image) -235.985 (and) -235.015 (a) -236.005 (ques\055) ] TJ T* [ (tion\054) -415.982 (it) -382.98 <02727374> -383.01 (generates) -382.992 (an) -382.995 (attention) -384.019 (map) -382.98 (on) -382.985 (image) -383.019 (re) 15.0073 (gions) ] TJ 11.884 TL T* [ (for) -365 (each) -363.986 (question) -365 (w) 10 (ord) -364.991 (and) -364.993 (an) -364.018 (attention) -365.003 (map) -364.983 (on) -364.988 (question) ] TJ 11.8832 TL T* [ (w) 10 (ords) -253.985 (for) -255.019 (each) -254.004 (image) -255.002 (re) 15.0098 (gion\056) -323.02 (It) -255.011 (then) -253.982 (performs) -255.002 (computation) ] TJ 11.884 TL T* [ (of) -226 (attended) -225.004 (features\054) -231.018 (concatenation) -224.997 (of) -226.001 (multimodal) -226.016 (represen\055) ] TJ 11.8828 TL T* [ (tations\054) -331.006 (and) -313.982 (their) -314.994 (transformation) -313.987 (by) -314.996 (a) -315.011 (single) -314.002 (layer) -315.006 (netw) 10.0081 (ork) ] TJ 11.884 TL T* [ (with) -453 (ReLU) -451.988 (and) -453.018 (a) -452.988 (residual) -453.003 (connection\056) -918.013 (These) -453.018 (computa\055) ] TJ 11.8832 TL T* [ (tions) -210.994 (are) -210.016 (encapsulated) -211.004 (into) -210.019 (a) -210.989 (composite) -210.004 (netw) 10.0081 (ork) -211.019 (that) -209.989 (we) -210.984 (call) ] TJ /R29 9.9626 Tf 11.884 TL T* [ (dense) -219.995 (co\055attention) ] TJ /R19 9.9626 Tf 75.225 0 Td [ (layer) 39.9933 (\054) -226 (since) -221.012 (it) -219.988 (considers) -220.997 (e) 25.0105 (v) 14.9828 (ery) -220.007 (interaction) ] TJ -75.225 -11.8828 Td [ (between) -196 (an) 15.0171 (y) -194.982 (image) -195.992 (re) 15.0098 (gion) -195.982 (and) -196.002 (an) 15.0171 (y) -194.982 (question) -196.011 (w) 10.0032 (ord\056) -291.981 (The) -196.002 (layer) ] TJ T* [ (has) -342.014 (fully) -342 (symmetric) -341.985 (architecture) -342.016 (between) -342.016 (the) -342.992 (tw) 10.0081 (o) -342.016 (modali\055) ] TJ 11.8832 TL T* [ (ties\054) -417.989 (and) -384.009 (can) -384.017 (be) -384.014 (stack) 10.0081 (ed) -384.014 (to) -384.005 (form) -384.005 (a) -384.019 (hierarch) 5.01877 (y) -384.009 (that) -384 (enables) ] TJ 11.884 TL T* [ (multi\055step) -249.988 (interactions) -250.002 (between) -249.993 (the) -249.988 (image\055question) -249.988 (pair) 55.0153 (\056) ] TJ 11.9547 -15.4398 Td [ (Starting) -390.994 (from) -390.984 (initial) -390.996 (representations) -391.003 (of) -390.994 (an) -392.013 (input) -391.003 (image) ] TJ -11.9547 -11.884 Td [ (and) -461.996 (question\054) -513.993 (each) -462.009 (dense) -461.992 (co\055attention) -462.006 (layer) -460.982 (in) -461.992 (the) -461.992 (layer) ] TJ 11.8832 TL T* [ (stack) -265.985 (updates) -266.995 (the) -265.987 (representations\054) -270 (which) -267 (are) -266.005 (inputted) -265.995 (to) -267.004 (the) ] TJ 11.884 TL T* [ (ne) 15.0171 (xt) -360.006 (layer) 55.0104 (\056) -638.989 (I) 1.01699 (ts) -360.006 <026e616c> -359.996 (output) -359.004 (are) -359.989 (then) -360.004 (fed) -359.004 (to) -360.009 (a) -359.004 (layer) -360.018 (for) -359.979 (an\055) ] TJ 11.8828 TL T* [ (swer) -226 (prediction\056) -301.996 (W) 79.9866 (e) -226.006 (use) -225.994 (additional) -226.001 (attention) -226.006 (mechanisms) -226.001 (in) ] TJ 11.884 TL T* [ (the) -211.016 (initial) -210.987 (feature) -212.011 (e) 15.0122 (xtraction) -211 (as) -210.984 (well) -211.009 (as) -210.984 (the) -211.994 (answer) -210.984 (prediction) ] TJ 11.8832 TL T* [ (layer) 55.0104 (\056) -529.007 (W) 79.9866 (e) -323.01 (call) -323 (the) -322.998 (entire) -322.998 (netw) 10.0081 (ork) -324.019 (including) -323 (all) -322.995 (these) -322.995 (com\055) ] TJ 11.884 TL T* [ (ponents) -422.998 (the) ] TJ /R29 9.9626 Tf 51.5949 0 Td [ (dense) -422.981 (coattention) -422.983 (network) ] TJ /R19 9.9626 Tf 112.257 0 Td [ (\050DCN\051\056) -423.013 (W) 79.9866 (e) -422.993 (sho) 24.9909 (w) ] TJ -163.852 -11.8828 Td [ (the) -301.981 (ef) 25.0081 (fecti) 25.0179 (v) 14.9828 (eness) -303 (of) -301.986 (DCNs) -302.984 (by) -302.018 (se) 25.0154 (v) 14.9828 (eral) -301.994 (e) 15.0122 (xperimental) -303.013 (results\073) ] TJ T* [ (the) 14.9852 (y) -324.019 (achie) 25.0154 (v) 14.9828 (e) -323.01 (the) -324.017 (ne) 25.0154 (w) -323.002 (state\055of\055the\055art) -324.015 (for) -323.005 (VQA) -324.015 (1\0560) -322.99 (and) -324.019 (2\0560) ] TJ 11.8832 TL (datasets\056) ' ET Q Q Q q q 1 1 1 rg /a0 gs 48.406 786.422 515.188 -52.699 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 14.4 TL ET 1 1 1 rg n 270 32 72 14 re f* 0.5 0.5 0.5 rg BT /F2 9 Tf 10.8 TL ET BT 1 0 0 1 297 35 Tm (6087) Tj T* ET Q endstream endobj 15 0 obj << /Filter /FlateDecode /Resources << /ExtGState << /a0 << /CA 1 /ca 1 >> >> /XObject << /x18 16 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 16 0 obj << /Filter /FlateDecode /Resources 17 0 R /Length 107 /Type /XObject /BBox [ 78 746 96 765 ] /Subtype /Form >> stream xe AC̬wʠ =p,?]%+H-
Jc "82w8VSnGW;"
endstream
endobj
17 0 obj
<<
/ExtGState <<
/a0 <<
/CA 1
/ca 1
>>
>>
>>
endobj
18 0 obj
<<
/Filter /FlateDecode
/Resources <<
/ExtGState <<
/a0 <<
/CA 1
/ca 1
>>
>>
/XObject <<
/x15 19 0 R
>>
>>
/Length 28
/Group <<
/Type /Group
/S /Transparency
/CS /DeviceRGB
/I true
>>
/BBox [ 67 752 84 775 ]
/Type /XObject
/Subtype /Form
>>
stream
x+O4PH/VЯ04Up
0
endstream
endobj
19 0 obj
<<
/Filter /FlateDecode
/Resources 20 0 R
/Length 228
/Type /XObject
/BBox [ 67 752 84 775 ]
/Subtype /Form
>>
stream
xeQKn!s ?FPav6R٪TS.
b];15YyR
{7QL.\:Rv/x9l+L7h%1!}i/AI(kz"U&,YO![R hg{3}4/GyYF:!w}Gn+'xJcO9i뽼_-:`
endstream
endobj
20 0 obj
<<
/ExtGState <<
/a0 <<
/CA 1
/ca 1
>>
>>
>>
endobj
21 0 obj
<<
/Filter /FlateDecode
/Resources <<
/ExtGState <<
/a0 <<
/CA 1
/ca 1
>>
>>
/XObject <<
/x24 22 0 R
>>
>>
/Length 28
/Group <<
/Type /Group
/S /Transparency
/CS /DeviceRGB
/I true
>>
/BBox [ 132 751 480 772 ]
/Type /XObject
/Subtype /Form
>>
stream
x+O4PH/VЯ02Qp
0
endstream
endobj
22 0 obj
<<
/Filter /FlateDecode
/Resources 23 0 R
/Length 53223
/Type /XObject
/BBox [ 132 751 480 772 ]
/Subtype /Form
>>
stream
xtI:6%Q㨈?7rA= u%6 ?Y(WbWo{B>9
x`Znϳ|8{3?0x*z ǃ|,@:w>`c|*ϻⳅKO3`g
:_|}}><.6`Z{{3]#<_o"~:ͺgk7/Ұ@|K yp ]03ʷCmş8˽Y?>(3!Bwqs.Z8,~~=rMT̩y+/*w: uBZ_`ߵp`%M?ɝ1ɳw=vDۉy&xb4Q>d@ sg~lA