%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 (Supervising Neural Attention Models for Video Captioning by Human Gaze Data) /Producer (PyPDF2) /Author (Youngjae Yu\054 Jongwook Choi\054 Yeonhwa Kim\054 Kyung Yoo\054 Sang\055Hun Lee\054 Gunhee Kim) /Subject (2017 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 /x6 17 0 R /x12 20 0 R /x10 23 0 R >> /ExtGState << /s9 26 0 R /s11 29 0 R /a0 << /CA 1 /ca 1 >> /R7 32 0 R /s5 33 0 R /s7 36 0 R >> /Font << /R10 39 0 R /F2 42 0 R /R12 43 0 R /F1 46 0 R /R14 47 0 R /R16 50 0 R /R8 53 0 R /R18 56 0 R /R20 59 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ ] >> endobj 4 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 63 0 R /Resources << /ExtGState << /R7 32 0 R >> /Font << /R10 39 0 R /R8 53 0 R /F2 64 0 R /R29 65 0 R /F1 68 0 R /R18 56 0 R /R35 69 0 R /R33 72 0 R /R27 75 0 R /R31 79 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ ] >> endobj 5 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 82 0 R /Resources << /ExtGState << /R7 32 0 R >> /Font << /R58 83 0 R /R48 88 0 R /R54 91 0 R /R46 96 0 R /R56 99 0 R /R44 104 0 R /R50 107 0 R /R42 110 0 R /R52 113 0 R /F2 117 0 R /F1 118 0 R /R10 39 0 R /R8 53 0 R /R29 65 0 R /R14 47 0 R /R60 119 0 R /R62 123 0 R /R18 56 0 R /R35 69 0 R /R27 75 0 R /R31 79 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ ] >> endobj 6 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 128 0 R /Resources << /XObject << /R105 129 0 R /R104 132 0 R /R127 133 0 R /R126 135 0 R /R125 134 0 R /R124 136 0 R /R123 137 0 R /R122 138 0 R /R121 140 0 R /R120 139 0 R /R83 141 0 R /R82 143 0 R /R81 144 0 R /R80 145 0 R /R85 147 0 R /R84 149 0 R /R163 151 0 R /R162 152 0 R /R161 153 0 R /R160 155 0 R /R167 157 0 R /R166 158 0 R /R165 159 0 R /R164 160 0 R /R169 161 0 R /R168 162 0 R /R78 163 0 R /R79 146 0 R /R72 165 0 R /R73 150 0 R /R71 142 0 R /R76 166 0 R /R77 164 0 R /R74 167 0 R /R75 148 0 R /R130 168 0 R /R131 169 0 R /R132 170 0 R /R133 171 0 R /R134 172 0 R /R135 173 0 R /R136 174 0 R /R137 175 0 R /R138 176 0 R /R139 177 0 R /R176 178 0 R /R170 179 0 R /R171 180 0 R /R114 181 0 R /R115 183 0 R /R109 184 0 R /R108 185 0 R /R147 187 0 R /R146 188 0 R /R141 189 0 R /R140 154 0 R /R143 190 0 R /R142 191 0 R /R101 192 0 R /R100 193 0 R /R103 194 0 R /R102 195 0 R /R149 196 0 R /R148 197 0 R /R107 186 0 R /R106 198 0 R /R129 199 0 R /R128 200 0 R /R145 201 0 R /R144 202 0 R /R90 203 0 R /R91 204 0 R /R156 205 0 R /R157 207 0 R /R118 208 0 R /R119 209 0 R /R152 210 0 R /R153 211 0 R /R150 212 0 R /R151 213 0 R /R112 214 0 R /R113 182 0 R /R110 216 0 R /R111 215 0 R /R116 217 0 R /R117 218 0 R /R158 219 0 R /R159 156 0 R /R154 220 0 R /R155 206 0 R >> /ExtGState << /R7 32 0 R /R89 130 0 R >> /Font << /R58 83 0 R /R48 88 0 R /R54 91 0 R /R46 96 0 R /R56 99 0 R /R44 104 0 R /R50 107 0 R /R42 110 0 R /R52 113 0 R /F2 221 0 R /F1 222 0 R /R92 223 0 R /R86 227 0 R /R97 230 0 R /R10 39 0 R /R8 53 0 R /R174 237 0 R /R177 241 0 R /R172 246 0 R /R95 249 0 R /R29 65 0 R /R62 123 0 R /R18 56 0 R /R35 69 0 R /R27 75 0 R /R31 79 0 R >> /ProcSet [ /Text /ImageC /ImageB /PDF /ImageI ] >> /Group 253 0 R /MediaBox [ 0 0 612 792 ] /Annots [ ] >> endobj 7 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 254 0 R /Resources << /ExtGState << /R7 32 0 R >> /Font << /R10 39 0 R /R58 83 0 R /R8 53 0 R /R48 88 0 R /R54 91 0 R /R46 96 0 R /R56 99 0 R /R44 104 0 R /R50 107 0 R /R42 110 0 R /R29 65 0 R /F1 255 0 R /R18 56 0 R /F2 256 0 R /R35 69 0 R /R27 75 0 R /R31 79 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ ] >> endobj 8 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 257 0 R /Resources << /XObject << /R192 258 0 R /R193 260 0 R /R190 259 0 R /R191 261 0 R /R196 262 0 R /R194 263 0 R /R195 264 0 R /R205 265 0 R /R206 266 0 R >> /ExtGState << /R7 32 0 R >> /Font << /R202 267 0 R /R10 39 0 R /F1 274 0 R /R14 47 0 R /R8 53 0 R /R198 275 0 R /F2 282 0 R /R27 75 0 R >> /ProcSet [ /Text /ImageC /ImageB /PDF /ImageI ] >> /Group 253 0 R /MediaBox [ 0 0 612 792 ] /Annots [ ] >> endobj 9 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 283 0 R /Resources << /ExtGState << /R7 32 0 R >> /Font << /R10 39 0 R /R213 284 0 R /R8 53 0 R /R54 91 0 R /R44 104 0 R /F2 287 0 R /R29 65 0 R /F1 288 0 R /R14 47 0 R /R35 69 0 R /R27 75 0 R /R31 79 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ ] >> endobj 10 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 289 0 R /Resources << /XObject << /R228 290 0 R /R229 291 0 R /R223 292 0 R /R236 293 0 R /R235 294 0 R /R234 295 0 R /R233 296 0 R /R225 297 0 R /R231 298 0 R /R227 299 0 R /R226 300 0 R /R232 301 0 R /R230 302 0 R /R222 303 0 R /R224 304 0 R >> /ExtGState << /R7 32 0 R >> /Font << /R10 39 0 R /F2 305 0 R /F1 306 0 R /R14 47 0 R /R8 53 0 R /R220 307 0 R /R86 227 0 R /R27 75 0 R >> /ProcSet [ /Text /ImageC /ImageB /PDF /ImageI ] >> /MediaBox [ 0 0 612 792 ] /Annots [ ] >> endobj 11 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 311 0 R /Resources << /ExtGState << /R7 32 0 R >> /Font << /F2 312 0 R /R8 53 0 R /R10 39 0 R /F1 313 0 R /R18 56 0 R /R27 75 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ ] >> endobj 12 0 obj << /Type /Catalog /Pages 1 0 R >> endobj 13 0 obj << /Length 20346 >> stream q q q 0.10000 0 0 0.10000 0 0 cm /R7 gs 0 0 0 rg q 10 0 0 10 0 0 cm BT /R8 14.34620 Tf 1 0 0 1 50.11210 675.06700 Tm [ (Super) 10.00260 (vising) -234.99800 (Neural) -236.99200 (Attention) -235.00700 (Models) -234.99700 (f) 24.99230 (or) -234.99200 (V) 37.01370 (ideo) -236.00800 (Captioning) -235.99200 (by) -235.00900 (Human) -235.99600 (Gaze) -235.00500 (Data) ] TJ /R10 11.95520 Tf 6.95391 -37.85780 Td [ (Y) 110.01500 (oungjae) -250.00800 (Y) 110.99500 (u) ] TJ /R12 7.97010 Tf 61.43710 4.33906 Td (\206) Tj /R10 11.95520 Tf 17.21290 -4.33906 Td [ (Jongw) 10.01360 (ook) -249.99000 (Choi) ] TJ /R12 7.97010 Tf 75.26990 4.33906 Td (\206) Tj /R10 11.95520 Tf 17.21290 -4.33906 Td [ (Y) 99.98470 (eonhw) 9.99010 (a) -249.99700 (Kim) ] TJ /R12 7.97010 Tf 68.74220 4.33906 Td (\207) Tj /R10 11.95520 Tf 17.21290 -4.33906 Td [ (K) 24.99770 (yung) -250.00200 (Y) 110.01600 (oo) ] TJ /R12 7.97010 Tf 54.50310 4.33906 Td (\207) Tj /R10 11.95520 Tf 17.21410 -4.33906 Td [ (Sang\055Hun) -250.00600 (Lee) ] TJ /R12 7.97010 Tf 69.38710 4.33906 Td (\207) Tj /R10 11.95520 Tf 17.21290 -4.33906 Td [ (Gunhee) -250.00600 (Kim) ] TJ /R12 7.97010 Tf 61.42500 4.33906 Td (\206) Tj -469.90800 -14.05820 Td (\206) Tj /R10 11.95520 Tf 4.26211 -4.33789 Td [ (Department) -250 (of) -250.01500 (Computer) -250.99400 (Science) -250.00600 (and) -249.98700 (Engineering) ] TJ /R12 7.97010 Tf 253.00100 4.33789 Td (\207) Tj /R10 11.95520 Tf 4.26172 -4.33789 Td [ (Department) -250 (of) -250.01400 (Brain) -251.00200 (\046) -249.99300 (Cogniti) 25.01200 (v) 14.98510 (e) -249.99700 (Sciences) ] TJ -123.69100 -13.94800 Td [ (Seoul) -249.99300 (National) -250 (Uni) 24.99570 (v) 14.98510 (ersity) 64.00850 (\054) -250.01200 (Seoul\054) -250.01400 (K) 34.99600 (orea) ] TJ /R14 8.96640 Tf -131.98100 -13.94800 Td (yj\056yu\100vision\056snu\056ac\056kr) Tj /R16 8.96640 Tf 128.32700 0 Td (\173) Tj /R14 8.96640 Tf 4.59961 0 Td [ (wookayin\054) -168.01000 (billygma\054) -167.00800 (yookyung20484\054) -167.00800 (visionsl\054) -167.00800 (gunhee) ] TJ /R16 8.96640 Tf 258.85100 0 Td (\175) Tj /R14 8.96640 Tf 4.59961 0 Td (\100snu\056ac\056kr) Tj /R8 11.95520 Tf -320.22300 -41.04570 Td (Abstract) Tj /R18 9.96260 Tf -83.92770 -23.91020 Td [ (The) -344.98600 (attention) -344.99600 (mec) 15.01100 (hanisms) -345.98400 (in) -344.98900 (deep) -346.01900 (neur) 14.99010 (al) -344.98900 (networks) -345.00600 (ar) 36.98650 (e) ] TJ -11.95510 -11.95510 Td [ (inspir) 37.01100 (ed) -319.00500 (by) -319.98500 (human\222) 40.01290 (s) -319 (attention) -321 (that) -318.98100 (sequentially) -319.00300 (focuses) -320.00500 (on) ] TJ 11.95590 TL T* [ (the) -255.98900 (most) -256.99500 (r) 37.01960 (ele) 15.01590 (vant) -255.98400 (parts) -258.00600 (of) -255.98400 (the) -257.00900 (information) -256.01100 (o) 10.00320 (ver) -257.01800 (time) -255.98400 (to) -257.00400 (g) 10.00320 (en\055) ] TJ 11.95510 TL T* [ (er) 15.01890 (ate) -330.99600 (pr) 36.98520 (ediction) -331.01500 (output\056) -333.00600 (The) -330.98900 (attention) -330.99900 (par) 14.99500 (ameter) 10.01060 (s) -332.01800 (in) -330.99100 (those) ] TJ T* [ (models) -308.00500 (ar) 36.98520 (e) -307.99200 (implicitly) -307.00500 (tr) 14.01110 (ained) -308.00900 (in) -308.01500 (an) -308.01700 (end\055to\055end) -308.01000 (manner) 110.98100 (\054) -307.98500 (yet) ] TJ T* [ (ther) 36.98830 (e) -323.00900 (have) -321.98400 (been) -323.00400 (fe) 14.98030 (w) -322.01100 (trials) -323.98500 (to) -322.99300 (e) 19.99180 (xplicitly) -321.99800 (incorpor) 15.01220 (ate) -322.99800 (human) ] TJ T* [ (gaze) -284.01900 (tr) 14.99140 (ac) 19.99730 (king) -284.01900 (to) -284.01900 (supervise) -284.98900 (the) -283.98400 (attention) -283.98700 (models\056) -284.00700 (In) -283.98900 (this) -284.00700 (pa\055) ] TJ T* [ (per) 111.00800 (\054) -444.98200 (we) -445.99500 (in) 40.00560 (vestigate) -445 (whether) -446.01500 (attention) -445.01900 (models) -445.00200 (can) -446 (bene) ] TJ /R20 9.96260 Tf 228.51200 0 Td <01> Tj /R18 9.96260 Tf 4.98125 0 Td (t) Tj -233.49300 -11.95590 Td [ (fr) 44.98640 (om) -330.99600 (e) 19.99240 (xplicit) -329.99800 (human) -330.98000 (gaze) -330.99100 (labels\054) -331.01500 (especially) -330.00400 (for) -330.98600 (the) -330.01600 (task) -330.98900 (of) ] TJ T* [ (video) -229.01200 (captioning) 14.98650 (\056) -229.99800 (W) 91.98590 (e) -230.00600 (collect) -229.98700 (a) -229.99600 (ne) 15.01830 (w) -229.00700 (dataset) -230.00800 (called) -229.00600 (V) 59.99490 (AS\054) -229.98900 (con\055) ] TJ T* [ (sisting) -279.00800 (of) -280.01900 (mo) 10.00630 (vie) -279.01300 (clips\054) -280.00800 (and) -280.01700 (corr) 36.98650 (esponding) -279.01200 (multiple) -279.99500 (descrip\055) ] TJ T* [ (tive) -420.98300 (sentences) -421.01600 (along) -421.98100 (with) -422.01400 (human) -421.00400 (gaze) -421.01600 (tr) 14.99010 (ac) 19.99670 (king) -421.01600 (data\056) -423.00300 (W) 91.98710 (e) ] TJ T* [ (pr) 44.98390 (opose) -371.02000 (a) -370.01200 (video) -370.98800 (captioning) -372.00800 (model) -370 (named) -370.98200 (Gaze) -370.98700 (Encoding) ] TJ 11.95630 TL T* [ (Attention) -400.00700 (Network) -400.01500 (\050GEAN\051) -401.99400 (that) -399.98700 (can) -400.00600 (le) 14.98030 (ver) 15.01470 (a) 10.00320 (g) 10.00320 (e) -400.01700 (gaze) -401.01900 (tr) 14.99010 (ac) 19.99670 (k\055) ] TJ 11.95510 TL T* [ (ing) -320.00800 (information) -320.00200 (to) -320.01300 (pr) 44.98390 (o) 10.00320 (vide) -321.00200 (the) -320.01800 (spatial) -321.00200 (and) -320.01000 (tempor) 15 (al) -320.01300 (atten\055) ] TJ T* [ (tion) -427.98300 (for) -427.00800 (sentence) -428 (g) 10.00320 (ener) 14 (ation\056) -428.01400 (Thr) 44.99860 (ough) -427.98700 (e) 15.01220 (valuation) -426.98500 (of) -427.99500 (lan\055) ] TJ T* [ (gua) 10.01360 (g) 10.00320 (e) -351.00500 (similarity) -350.98800 (metrics) -351.99900 (and) -350.98600 (human) -351.99600 (assessment) -351.00300 (via) -350.99300 (Ama\055) ] TJ T* [ (zon) -427.99100 (mec) 15.01100 (hanical) -426.99600 (T) 54.98590 (urk\054) -429.01700 (we) -427.99800 (demonstr) 15.01100 (ate) -427.01900 (that) -427.98300 (spatial) -428.00500 (atten\055) ] TJ T* [ (tions) -204.99400 (guided) -205.00200 (by) -206.00400 (human) -205.98200 (gaze) -205.01200 (data) -205.00300 (indeed) -205.01000 (impr) 44.99490 (o) 10.00320 (ve) -206.01500 (the) -205.01700 (perfor) 19.98200 (\055) ] TJ 11.95590 TL T* [ (mance) -283.98800 (of) -283 (multiple) -283.01400 (captioning) -285.00300 (methods\056) -284.00200 (Mor) 36.98650 (eo) 9.99833 (ver) 110.99900 (\054) -283.00900 (we) -283.98200 (show) ] TJ 11.95510 TL T* [ (that) -466.01500 (the) -465.99300 (pr) 44.98390 (oposed) -466.01800 (appr) 44.01350 (oac) 14.98400 (h) -465.99600 (ac) 15.01830 (hie) 14.98520 (ves) -466.00300 (the) -465.99300 (state\055of\055the\055art) ] TJ T* [ (performance) -258.00600 (for) -258.01700 (both) -258.99400 (gaze) -259.00300 (pr) 36.98520 (ediction) -258.00700 (and) -257.98100 (video) -259.00800 (captioning) ] TJ T* [ (not) -350.00300 (only) -348.98800 (in) -350.00800 (our) -348.98500 (V) 59.99370 (AS) -350.01100 (dataset) -349.98900 (b) 20.00160 (ut) -350.00800 (also) -349.99700 (in) -348.98800 (standar) 37.01340 (d) -350.01500 (datasets) ] TJ (\050) ' /R10 9.96260 Tf 3.31758 0 Td (e\056g) Tj /R18 9.96260 Tf 11.89530 0 Td [ (\056) -249.99500 (LSMDC) -250 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R18 9.96260 Tf 1 0 0 1 109.33000 255.33100 Tm (24) Tj ET Q 0 0 0 rg q 10 0 0 10 0 0 cm BT /R18 9.96260 Tf 1 0 0 1 119.29300 255.33100 Tm [ (\135) -249.99100 (and) -249.98300 (Hollywood2) -249.98500 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R18 9.96260 Tf 1 0 0 1 198.16600 255.33100 Tm (18) Tj ET Q 0 0 0 rg q 10 0 0 10 0 0 cm BT /R18 9.96260 Tf 1 0 0 1 208.12900 255.33100 Tm (\135\051\056) Tj /R8 11.95520 Tf -158.01700 -35.85000 Td [ (1\056) -249.99000 (Intr) 18.01460 (oduction) ] TJ /R10 9.96260 Tf 11.95510 -18.92890 Td [ (Attention\055based) -360.99100 (models) -363.00700 (ha) 19.99790 (v) 14.98280 (e) -361.00300 (recently) -362.01100 (g) 4.98446 (ained) -360.99100 (much) -362 (in\055) ] TJ -11.95510 -11.95590 Td [ (terest) -441.99600 (as) -440.98300 (a) -442.00900 (po) 24.98600 (werful) -442 (deep) -442.98400 (neural) -441.02000 (netw) 10.00810 (ork) -441.99700 (architecture) -442.00200 (in) ] TJ 11.95510 TL T* [ (a) -319.99000 (v) 24.98110 (ariety) -320.01800 (of) -321.00300 (applications\054) -321 (including) -319.98300 (image) -320.01000 (captioning) -321.00800 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R10 9.96260 Tf 1 0 0 1 270.59400 176.64100 Tm (35) Tj ET Q 0 0 0 rg q 10 0 0 10 0 0 cm BT /R10 9.96260 Tf 1 0 0 1 280.55700 176.64100 Tm (\135\054) Tj -230.44500 -11.95510 Td [ (video) -205.98700 (captioning) -206.00600 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R10 9.96260 Tf 1 0 0 1 121.73300 164.68600 Tm (15) Tj ET Q 0 0 0 rg q 10 0 0 10 0 0 cm BT /R10 9.96260 Tf 1 0 0 1 131.69600 164.68600 Tm [ (\135\054) -204.98700 (action) -205.98900 (recognition) -206.01000 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R10 9.96260 Tf 1 0 0 1 216.69600 164.68600 Tm (27) Tj ET Q 0 0 0 rg q 10 0 0 10 0 0 cm BT /R10 9.96260 Tf 1 0 0 1 226.65900 164.68600 Tm [ (\135\054) -206.00500 (object) -205.99000 (recog\055) ] TJ -176.54700 -11.95510 Td [ (nition) -313.98900 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R10 9.96260 Tf 1 0 0 1 79.81050 152.73100 Tm (1) Tj ET Q 0 0 0 rg q 10 0 0 10 0 0 cm BT /R10 9.96260 Tf 1 0 0 1 84.79180 152.73100 Tm [ (\135\054) -313.98700 (and) -313.98100 (machine) -314.99000 (translation) -315.00700 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R10 9.96260 Tf 1 0 0 1 196.64200 152.73100 Tm (2) Tj ET Q 0 0 0 rg q 10 0 0 10 0 0 cm BT /R10 9.96260 Tf 1 0 0 1 201.62300 152.73100 Tm [ (\135) -313.98900 (to) -314.01400 (name) -314.01900 (a) -313.99200 (fe) 25.00810 (w) 65.01120 (\056) -315.00400 (The) ] TJ -151.51100 -11.95470 Td [ (attention) -442.98900 (models) -444.01300 (are) -442.99700 (loosely) -444.99300 (inspired) -442.99600 (by) -443.01400 (visual) -444.01700 (attention) ] TJ 11.95630 TL T* [ (mechanism) -273.99800 (of) -273.99200 (humans\054) -274.98100 (who) -273.98100 (do) -273.98400 (not) -274.01600 (focus) -273.98600 (their) -273.98100 (attention) -273.99800 (on) ] TJ 11.95510 TL T* [ (the) -289.00300 (entire) -287.98400 (scene) -289.01300 (at) -289.00800 (once\054) -289.98700 (b) 20.00160 (ut) -288.99800 (instead) -288.00800 (sequentially) -289.00800 (adjust) -288.98400 (the) ] TJ T* [ (focal) -249.99700 (points) -250.00100 (on) -249.98800 (dif) 24.98480 (ferent) -250.98000 (parts) -249.98600 (of) -249.99600 (the) -249.99000 (scene) -250 (o) 14.98280 (v) 14.98280 (er) -250.00500 (time\056) ] TJ 11.95510 -11.95510 Td [ (Although) -344.99800 (the) -344.99400 (attention) -345.98500 (models) -345.01000 (simulate) -345.00400 (human\222) 54.99080 (s) -344.99600 (atten\055) ] TJ -11.95510 -11.95510 Td [ (tion\054) -252.00900 (surprisingly) -251.00300 (there) -253.01700 (ha) 19.99790 (v) 14.98280 (e) -250.98300 (been) -251.99600 (fe) 25.00810 (w) -251.01400 (trials) -252.01700 (to) -251.00500 (e) 15.01220 (xplicitly) -251.99000 (in\055) ] TJ 258.75000 473.21000 Td [ (corporate) -216 (human) -215 (g) 4.98446 (aze) -216.99600 (tracking) -215.00300 (labels) -215.98800 (to) -215.99300 (supervise) -214.99800 (the) -215.99300 (atten\055) ] TJ T* [ (tion) -266.99000 (mechanism\056) -265.99700 (Usually) -268.00400 (attention) -266 (models) -266.98500 (are) -266.98500 (trained) -265.98000 (in) -267.00400 (an) ] TJ T* [ (end\055to\055end) -191.01000 (manner) 40.00310 (\054) -192.00500 (and) -192.00200 (thus) -191.98300 (attention) -190.99300 (weights) -190.99800 (are) -192.01700 (implicitly) ] TJ 11.95590 TL T* [ (learned\056) -319.98300 (In) -319.00300 (this) -319.02000 (paper) 39.99090 (\054) -319.98300 (we) -319.98600 (aim) -318.99100 (at) -319 (in) 40.00310 (v) 14.98280 (estig) 5 (ating) -318.98600 (whether) -320.01500 (the) ] TJ 11.95510 TL T* [ (e) 15.01220 (xplicit) -196.01100 (human) -197.00400 (g) 4.98446 (aze) -196.01900 (labels) -197.01100 (can) -196.99200 (better) -195.99700 (guide) -195.98700 (attention) -196.99200 (models) ] TJ T* [ (and) -394.98800 (e) 25.01050 (v) 14.98280 (entually) -393.98600 (enhance) -396.01500 (their) -394.98300 (prediction) -393.98300 (performance\056) -396.00700 (W) 79.98660 (e) ] TJ T* [ (focus) -255.98900 (on) -255.98700 (the) -257.00900 (task) -255.98200 (of) ] TJ /R18 9.96260 Tf 80.82620 0 Td [ (video) -255.98700 (captioning) ] TJ /R10 9.96260 Tf 66.74920 0 Td [ (\054) -257.01100 (whose) -255.98200 (objecti) 25 (v) 14.98280 (e) -256.98200 (is) -255.98200 (to) ] TJ -147.57500 -11.95510 Td [ (generate) -335.00500 (a) -335.98800 (descripti) 25 (v) 14.98280 (e) -335.98800 (sentence) -336.01700 (for) -335.00300 (a) -335.98800 (gi) 24.98600 (v) 14.98280 (en) -335.00300 (video) -336.00300 (clip\056) -336.00300 (W) 79.98660 (e) ] TJ T* [ (choose) -348.98800 (the) -348.01300 (video) -348.98300 (captioning) -349.98100 (because) -347.98100 (the) -348.99100 (attention) -349.00500 (mecha\055) ] TJ 11.95630 TL T* [ (nism) -280.00200 (may) -281 (ha) 19.99670 (v) 14.98280 (e) -279.99700 (more) -281.98200 (room) -280.01200 (to) -281.00200 (play) -280.01700 (a) -279.99700 (role) -281.00700 (in) -280.02200 (summarizing) -281.00200 (a) ] TJ 11.95470 TL T* [ (sequence) -263.01000 (of) -263.99300 (frames) -263.01200 (that) -264.02000 (may) -263.98000 (contain) -262.99500 (too) -262.99500 (much) -264.01500 (information) ] TJ T* [ (for) -214.98600 (a) -214.98900 (short) -215.00300 (output) -214.98900 (sentence\056) -215.99300 (It) -215.01800 (is) -215.98800 (w) 10.00320 (orth) -215.00800 (noting) -214.98900 (that) -215.00800 (our) -215.01800 (objec\055) ] TJ T* [ (ti) 24.99090 (v) 14.98280 (e) -277.01700 (is) -277 (not) -276.99500 (to) -277 (replace) -276.98100 (e) 15.01220 (xisting) -277.98800 (video) -276.99300 (captioning) -277.01300 (methods) -277.01700 (for) ] TJ T* [ (e) 25.01050 (v) 14.98280 (ery) -345.98400 (use) -344.99400 (case\054) -345.98600 (gi) 24.98850 (v) 14.98280 (en) -345.00100 (that) -346.98600 (acquisition) -344.99600 (of) -346.02100 (human) -345.01600 (g) 4.98446 (aze) -345.99600 (data) ] TJ 11.95590 TL T* [ (is) -274.02000 (e) 15.01220 (xpensi) 25 (v) 14.98280 (e\056) -273.99600 (Instead\054) -274.99800 (we) -273.99600 (study) -272.98400 (the) -273.98400 (ef) 25.00560 (fect) -273.99800 (of) -273.99300 (supervision) -273.98900 (by) ] TJ 11.95510 TL T* [ (human) -343.99900 (g) 4.98446 (aze) -343.99400 (for) -343.00400 (attention) -345.00600 (mechanism\054) -343.98200 (which) -344.00600 (has) -344.01600 (not) -344.00100 (been) ] TJ T* [ (discussed) -250.00500 (in) -249.98500 (pre) 25.01300 (vious) -250.00700 (literature\056) ] TJ 11.95510 -14.27500 Td [ (W) 79.98660 (e) -347.00600 (collect) -346.98600 (a) -347.00600 (ne) 25.01540 (w) -346.99800 (dataset) -348.98600 (named) -346.98600 (V) 135.00200 (AS) -346.98100 (\050) ] TJ /R18 9.96260 Tf 160.70600 0 Td [ (V) 73.99260 (isual) -347.01100 (Attentive) ] TJ -172.66100 -11.95470 Td (Script) Tj /R10 9.96260 Tf 23.98010 0 Td [ (\051\054) -464.00100 (consisting) -463.99100 (of) -463.99900 (mo) 14.01730 (vie) -463.99100 (videos) -464.01600 (of) -464.00100 (15) -463.99100 (seconds) -463.98100 (long\054) ] TJ -23.98010 -11.95510 Td [ (with) -280.01000 (multiple) -279.00500 (descripti) 23.98130 (v) 14.98280 (e) -279.99700 (sentences) -279.98300 (and) -279.00700 (g) 4.98446 (aze) -280.00700 (tracking) -280.01200 (data\056) ] TJ T* [ (F) 14.99260 (or) -420.98600 (pretraining) -421.98100 (and) -420.98400 (e) 25.01050 (v) 24 (aluation) -420.99800 (of) -422.00800 (models\054) -420.99300 (we) -420.98900 (also) -421.99300 (le) 25.02030 (v) 14.98280 (er) 19.98690 (\055) ] TJ T* [ (age) -208.98700 (lar) 17.99700 (ge\055scale) -210.00700 (caption\055only) -210.00400 (LSMDC) -210 (dataset) -209.00900 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R10 9.96260 Tf 1 0 0 1 491.75500 300.83200 Tm (24) Tj ET Q 0 0 0 rg q 10 0 0 10 0 0 cm BT /R10 9.96260 Tf 1 0 0 1 501.71800 300.83200 Tm [ (\135) -208.99000 (and) -210 (g) 4.98446 (aze\055) ] TJ -192.85600 -11.95510 Td [ (only) -250.01500 (Hollyw) 9.98853 (ood2) -250.01700 (e) 15.01220 (ye) -251.01700 (mo) 14.99750 (v) 14.98280 (ement) -249.98300 (dataset) -249.98300 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R10 9.96260 Tf 1 0 0 1 475.16800 288.87700 Tm (17) Tj ET Q 0 0 0 rg q 10 0 0 10 0 0 cm BT /R10 9.96260 Tf 1 0 0 1 485.13000 288.87700 Tm (\054) Tj ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R10 9.96260 Tf 1 0 0 1 490.11200 288.87700 Tm (18) Tj ET Q 0 0 0 rg q 10 0 0 10 0 0 cm BT /R10 9.96260 Tf 1 0 0 1 500.07400 288.87700 Tm (\135\056) Tj -179.25700 -14.27500 Td [ (T) 79.99160 (o) -215.99800 (e) 15.01220 (xplicitly) -215.99600 (model) -216.99800 (the) -217.01300 (g) 4.98446 (aze) -216.01800 (prediction) -216.01300 (for) -216.98300 (sentence) -215.99800 (gen\055) ] TJ -11.95510 -11.95510 Td [ (eration\054) -264.02000 (we) -263.01700 (propose) -264.01500 (a) -263.02000 (no) 14.00750 (v) 14.98280 (el) -263.99000 (video) -264.01500 (captioning) -263.01500 (model) -264.01000 (named) ] TJ /R18 9.96260 Tf 11.95470 TL T* [ (Gaze) -390.00400 (Encoding) -389.98100 (Attention) -390.98900 (Network) ] TJ /R10 9.96260 Tf 144.67700 0 Td [ (\050GEAN\051\056) -390.00400 (The) -390.00800 (encoder) ] TJ -144.67700 -11.95630 Td [ (generates) -394.99000 (pools) -395.00300 (of) -396.01000 (visual) -395.98800 (features) -394.99300 (depending) -395.00300 (on) -394.98300 (not) -395.99300 (only) ] TJ T* [ (content) -258.01600 (and) -259.01100 (motion) -257.99400 (in) -259.00300 (videos\054) -258.99100 (b) 20.00160 (ut) -257.98600 (also) -257.98100 (g) 4.98446 (aze) -258.01100 (maps) -258.99600 (predicted) ] TJ T* [ (by) -239.98900 (the) -239.99200 (recurrent) -240.00400 (g) 4.98446 (aze) -240.01100 (prediction) -240.98900 (\050RGP\051) -240.01400 (model\056) -240.00900 (The) -239.99400 (decoder) ] TJ T* [ (generates) -317.98400 (w) 10 (ord) -318.99800 (sequences) -318.99600 (by) -318.99600 (dynamically) -317.98600 (focusing) -318 (on) -318.99600 (the) ] TJ T* [ (most) -250.00700 (rele) 25.01050 (v) 24.98110 (ant) -249.99000 (subsets) -250.98500 (of) -249.99500 (the) -249.98800 (feature) -250.00200 (pools\056) ] TJ 11.95510 -14.27500 Td [ (Through) -532.00400 (quantitati) 25.01790 (v) 14.98280 (e) -533.99400 (e) 25.01050 (v) 24.98110 (aluation) -532 (using) -532 (language) -532.99400 (met\055) ] TJ -11.95510 -11.95510 Td [ (rics) -354.99500 (and) -354.99500 (human) -355.01700 (assessment) -355.98500 (via) -354.99000 (Amazon) -356.00400 (Mechanical) -354.98500 (T) 45.01700 (urk) ] TJ T* [ (\050AMT\051\054) -221.98700 (we) -222.00500 (sho) 24.99340 (w) -223.01900 (that) -223.00700 (human) -222.01700 (g) 4.98446 (aze) -222.01700 (indeed) -221.98700 (helps) -223.00200 (enhance) -222.00200 (the) ] TJ 11.95590 TL T* [ (video) -211.98700 (captioning) -212.00600 (accurac) 14.01730 (y) -211.01900 (of) -212.00400 (attention) -212.00900 (models\056) -212.00900 (One) -211.99900 (promis\055) ] TJ 11.95510 TL T* [ (ing) -352.98300 (result) -354.00200 (is) -352.98800 (that) -352.98500 (our) -354.01200 (model) -353.99500 (learns) -354.01400 (from) -352.99000 (a) -353.00500 (relati) 24.98600 (v) 14.98280 (ely) -354.01000 (small) ] TJ T* [ (amount) -266.00500 (of) -265.99200 (g) 4.98446 (aze) -266.98700 (data) -265.99200 (of) -267.01200 (V) 135.00200 (AS) -266.01400 (and) -265.99000 (Hollyw) 9.98853 (ood2) -266.99500 (datasets\054) -266.01400 (and) ] TJ T* [ (impro) 15.00480 (v) 14.98280 (es) -225.01900 (the) -225.01400 (captioning) -226.00400 (quality) -224.98700 (on) -225.01100 (LSMDC) -225.01600 (dataset) -225.00700 (with) -224.99700 (no) ] TJ T* [ (g) 4.98446 (aze) -292.98300 (annotation\056) -293.99000 (It) -293.00500 (hints) -294 (that) -292.99500 (potentially) -293 (we) -293.99000 (could) -292.99000 (le) 25.02030 (v) 14.98280 (erage) ] TJ T* [ (g) 4.98446 (aze) -367.01000 (information) -366.99300 (in) -366.98500 (a) -367.98300 (semi\055supervised) -366.99300 (manner) 40.00310 (\054) -366.99300 (and) -366.99300 (apply) ] 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 1 299.25000 35 Tm (490) 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 /FlateDecode
/Resources <<
/ExtGState <<
/a0 <<
/CA 1
/ca 1
>>
>>
/XObject <<
/x15 18 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
18 0 obj
<<
/Filter /FlateDecode
/Resources 19 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
19 0 obj
<<
/ExtGState <<
/a0 <<
/CA 1
/ca 1
>>
>>
>>
endobj
20 0 obj
<<
/Filter /FlateDecode
/Resources <<
/ExtGState <<
/a0 <<
/CA 1
/ca 1
>>
>>
/XObject <<
/x24 21 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
21 0 obj
<<
/Filter /FlateDecode
/Resources 22 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