%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 (Improved Image Captioning via Policy Gradient Optimization of SPIDEr) /Producer (PyPDF2) /Author (Siqi Liu\054 Zhenhai Zhu\054 Ning Ye\054 Sergio Guadarrama\054 Kevin Murphy) /Subject (2017 IEEE International Conference on Computer Vision) >> 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 >> /R19 32 0 R /s5 33 0 R /s7 36 0 R >> /Font << /F2 39 0 R /R28 40 0 R /F1 43 0 R /R20 44 0 R /R22 49 0 R /R32 53 0 R /R24 57 0 R /R30 61 0 R /R26 65 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 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 ] >> endobj 4 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 81 0 R /Resources << /ExtGState << /R19 32 0 R >> /Font << /F2 82 0 R /R59 83 0 R /R73 87 0 R /R71 90 0 R /R75 93 0 R /R65 97 0 R /R67 101 0 R /F1 106 0 R /R61 107 0 R /R63 111 0 R /R20 44 0 R /R22 49 0 R /R69 114 0 R /R26 65 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 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 ] >> endobj 5 0 obj << /Parent 1 0 R /Rotate 0 /Contents 135 0 R /Resources << /XObject << /R95 136 0 R /R96 137 0 R >> /ExtGState << /R19 32 0 R >> /Font << /F2 138 0 R /R73 87 0 R /R71 90 0 R /R75 93 0 R /R65 97 0 R /R28 40 0 R /R67 101 0 R /F1 139 0 R /R61 107 0 R /R63 111 0 R /R26 65 0 R /R20 44 0 R /R22 49 0 R /R69 114 0 R /R97 140 0 R >> /ProcSet [ /Text /ImageC /ImageB /PDF /ImageI ] >> /Group 143 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 144 0 R 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 ] /Type /Page >> endobj 6 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 155 0 R /Resources << /ExtGState << /R19 32 0 R >> /Font << /F2 156 0 R /R73 87 0 R /R71 90 0 R /R114 157 0 R /R75 93 0 R /R65 97 0 R /R28 40 0 R /R67 101 0 R /F1 160 0 R /R61 107 0 R /R63 111 0 R /R26 65 0 R /R20 44 0 R /R22 49 0 R /R69 114 0 R /R97 140 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 161 0 R 162 0 R 163 0 R 164 0 R 165 0 R 166 0 R 167 0 R 168 0 R 169 0 R ] >> endobj 7 0 obj << /Parent 1 0 R /Rotate 0 /Contents 170 0 R /Resources << /XObject << /R138 171 0 R /R139 172 0 R >> /ExtGState << /R19 32 0 R >> /Font << /F2 173 0 R /R75 93 0 R /R65 97 0 R /R28 40 0 R /F1 174 0 R /R61 107 0 R /R63 111 0 R /R20 44 0 R /R22 49 0 R /R69 114 0 R >> /ProcSet [ /Text /ImageC /ImageB /PDF /ImageI ] >> /Group 143 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 175 0 R 176 0 R 177 0 R 178 0 R 179 0 R 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 ] /Type /Page >> endobj 8 0 obj << /Parent 1 0 R /Rotate 0 /Contents 192 0 R /Resources << /XObject << /R156 193 0 R /R157 194 0 R >> /ExtGState << /R19 32 0 R >> /Font << /F2 195 0 R /R20 44 0 R /F1 196 0 R /R22 49 0 R >> /ProcSet [ /Text /ImageC /ImageB /PDF /ImageI ] >> /Group 143 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 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 ] /Type /Page >> endobj 9 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 207 0 R /Resources << /XObject << /R168 208 0 R /R167 209 0 R /R166 210 0 R /R165 211 0 R >> /ExtGState << /R19 32 0 R >> /Font << /F2 212 0 R /R20 44 0 R /F1 213 0 R /R22 49 0 R /R61 107 0 R >> /ProcSet [ /Text /ImageC /ImageB /PDF /ImageI ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 214 0 R 215 0 R ] >> endobj 10 0 obj << /Parent 1 0 R /Rotate 0 /Contents 216 0 R /Resources << /XObject << /R190 217 0 R /R185 219 0 R /R184 220 0 R /R187 222 0 R /R186 223 0 R /R189 218 0 R /R188 224 0 R /R183 221 0 R >> /ExtGState << /R19 32 0 R >> /Font << /F2 225 0 R /R20 44 0 R /F1 226 0 R /R22 49 0 R /R28 40 0 R >> /ProcSet [ /Text /ImageC /ImageB /PDF /ImageI ] >> /Group 143 0 R /MediaBox [ 0 0 612 792 ] /Annots [ 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 ] /Type /Page >> endobj 11 0 obj << /Parent 1 0 R /Rotate 0 /Type /Page /Contents 236 0 R /Resources << /ExtGState << /R19 32 0 R >> /Font << /F2 237 0 R /R20 44 0 R /F1 238 0 R /R22 49 0 R >> /ProcSet [ /ImageC /Text /PDF /ImageI /ImageB ] >> /MediaBox [ 0 0 612 792 ] /Annots [ 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 248 0 R 249 0 R 250 0 R 251 0 R 252 0 R 253 0 R 254 0 R 255 0 R 256 0 R 257 0 R 258 0 R 259 0 R 260 0 R 261 0 R 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 ] >> endobj 12 0 obj << /Type /Catalog /Pages 1 0 R >> endobj 13 0 obj << /Length 18406 >> stream q q q 0.10000 0 0 0.10000 0 0 cm /R19 gs 0 g q 10 0 0 10 0 0 cm BT /R20 14.34620 Tf 1 0 0 1 75.26210 675.06700 Tm [ (Impr) 17.99210 (o) 9.99625 (v) 9.99625 (ed) -250.00400 (Image) -250.00500 (Captioning) -249.98800 (via) -250.00700 (P) 20.00610 (olicy) -249.99600 (Gradient) -249.99800 (optimization) -249.98800 (of) -250.01200 (SPIDEr) ] TJ /R22 11.95520 Tf 35.15780 -47.82030 Td [ (Siqi) -249.99700 (Liu) ] TJ /R24 7.97010 Tf 38.86600 4.33828 Td [ <03> -0.29866 ] TJ /R22 7.97010 Tf 4.73203 0 Td (1) Tj /R22 11.95520 Tf 4.48320 -4.33828 Td [ (\054) -250.01200 (Zhenhai) -250.00300 (Zhu) ] TJ /R22 7.97010 Tf 67.40270 4.33828 Td (2) Tj /R22 11.95520 Tf 4.48320 -4.33828 Td [ (\054) -250.01200 (Ning) -250.00200 (Y) 99.98470 (e) ] TJ /R22 7.97010 Tf 45.62110 4.33828 Td (2) Tj /R22 11.95520 Tf 4.48281 -4.33828 Td [ (\054) -250.01200 (Ser) 18 (gio) -250.00400 (Guadarrama) ] TJ /R22 7.97010 Tf 99.04800 4.33828 Td (2) Tj /R22 11.95520 Tf 4.48281 -4.33828 Td [ (\054) -250.01200 (and) -249.98700 (K) 24.99570 (e) 24.98750 (vin) -250.00200 (Murph) 4.98688 (y) ] TJ /R22 7.97010 Tf 96.29920 4.33828 Td (2) Tj /R26 11.95520 Tf -352.42300 -28.24920 Td (f) Tj /R28 11.95520 Tf 5.97813 0 Td (liusiqi\054zhenhai\054nye\054sguada\054kpmurphy) Tj /R26 11.95520 Tf 251.05800 0 Td (g) Tj /R28 11.95520 Tf 5.97813 0 Td (\100google\056com) Tj /R22 7.97010 Tf -228.94200 -9.60898 Td (1) Tj /R22 11.95520 Tf 4.48320 -4.33789 Td [ (Department) -250 (of) -250.01400 (Computer) -250.01400 (Science\054) -249.99300 (Uni) 24.99570 (v) 14.98510 (ersity) -249.98900 (of) -250.01400 (Oxford) ] TJ /R22 7.97010 Tf 111.31900 -9.61016 Td (2) Tj /R22 11.95520 Tf 4.48438 -4.33789 Td (Google) Tj /R20 11.95520 Tf -136.26100 -49.41290 Td (Abstract) Tj /R30 9.96260 Tf -83.92770 -24.09800 Td [ (Curr) 37.00490 (ent) -382.00700 (ima) 10.01300 (g) 10.00320 (e) -381 (captioning) -382.00600 (methods) -381.00800 (ar) 36.98650 (e) -382.02000 (usually) -381.98300 (tr) 14.99010 (ained) ] TJ -11.95510 -11.95510 Td [ (via) -792.01700 (maximum) -792.00400 (lik) 10.00810 (elihood) -791.99400 (estimation\056) -1937 (Howe) 14.99500 (ver) 110.99900 (\054) -928 (the) ] TJ 11.95510 TL T* [ (lo) 10.01060 (g\055lik) 10.01000 (elihood) -414.01800 (scor) 36.98650 (e) -412.99500 (of) -413.99700 (a) -412.98500 (caption) -414.01800 (does) -413.99700 (not) -413.01200 (corr) 36.98650 (elate) -414 (well) ] TJ T* [ (with) -492.00300 (human) -492.99300 (assessments) -491.99400 (of) -493.00400 (quality) 55.00310 (\056) -1037 (Standar) 37.01340 (d) -493.01100 (syntactic) ] TJ T* [ (e) 15.01280 (valuation) -261.99200 (metrics\054) -264.99100 (suc) 14.98520 (h) -263.01000 (as) -261.98400 (BLEU) 25 (\054) -261.99300 (METEOR) -262.00300 (and) -263 (R) 39.99820 (OUGE\054) ] TJ 11.95630 TL T* [ (ar) 36.98520 (e) -271.99800 (also) -272.01000 (not) -272.01600 (well) -272.99000 (corr) 36.98650 (elated\056) -375.99700 (The) -272.01800 (ne) 15.01830 (wer) -272.01800 (SPICE) -272.99100 (and) -272.01800 (CIDEr) ] TJ 11.95470 TL T* [ (metrics) -387.01300 (ar) 36.98520 (e) -386.99900 (better) -385.99700 (corr) 36.98650 (elated\054) -421.00900 (b) 20.00160 (ut) -386.98200 (have) -386.99200 (tr) 14.99010 (aditionally) -387.01600 (been) ] TJ T* [ (har) 36.99020 (d) -533.00400 (to) -532.99700 (optimize) -533.01300 (for) 111.01900 (\056) -1158.98000 (In) -533.00800 (this) -532.98300 (paper) 111.01800 (\054) -604.01400 (we) -533 (show) -533.01900 (how) -532.98700 (to) ] TJ T* [ (use) -382.00800 (a) -380.99000 (policy) -381.99900 (gr) 14.98890 (adient) -381.99000 (\050PG\051) -380.99500 (method) -381.99400 (to) -380.98300 (dir) 36.99140 (ectly) -382.01200 (optimize) -382.01700 (a) ] TJ T* [ (linear) -395.98600 (combination) -394.98200 (of) -396 (SPICE) -395.98900 (and) -395.01700 (CIDEr) -395.99000 (\050a) -396.01000 (combination) ] TJ T* [ (we) -398.98300 (call) -399.01600 (SPIDEr\051\072) -608.00400 (the) -398.98500 (SPICE) -399.00800 (scor) 36.98650 (e) -400.01700 (ensur) 38.01080 (es) -400.00900 (our) -399.01400 (captions) ] TJ 11.95630 TL T* [ (ar) 36.98520 (e) -345.98600 (semantically) -345.98100 (faithful) -344.98900 (to) -346.00800 (the) -346.01300 (ima) 10.01300 (g) 10.00320 (e) 9.99343 (\054) -370.01400 (while) -345.99100 (CIDEr) -345.99900 (scor) 36.98650 (e) ] TJ 11.95510 TL T* [ (ensur) 36.99200 (es) -521.99000 (our) -522.99500 (captions) -521.99500 (ar) 36.98520 (e) -523.01600 (syntactically) -521.99100 <0375656e742e> -1127.01000 (The) -522.99600 (PG) ] TJ T* [ (method) -195.98600 (we) -197.01700 (pr) 44.98390 (opose) -195.98900 (impr) 44.99370 (o) 10.00320 (ves) -196.00900 (on) -197.01600 (the) -196 (prior) -196.01600 (MIXER) -197.01600 (appr) 44.99490 (oac) 14.98280 (h\054) ] TJ T* [ (by) -531.98900 (using) -532 (Monte) -532.99400 (Carlo) -531.99200 (r) 45.01820 (ollouts) -531.98600 (instead) -531.99700 (of) -532.01700 (mixing) -532.98700 (MLE) ] TJ T* [ (tr) 14.99140 (aining) -313.99200 (with) -313.99400 (PG\056) -314.99600 (W) 91.98590 (e) -313.99200 (show) -313.99700 (empirically) -313.99200 (that) -314.00200 (our) -314.98900 (algorithm) ] TJ T* [ (leads) -223.98300 (to) -223.00900 (easier) -223.98600 (optimization) -224.01000 (and) -223.00700 (impr) 44.99370 (o) 10.00320 (ved) -224.00400 (r) 37.01830 (esults) -224.01400 (compar) 36.99380 (ed) ] TJ 11.95590 TL T* [ (to) -296.99700 (MIXER\056) -296.99600 (F) 45.01580 (inally) 54.99820 (\054) -307.98500 (we) -297 (show) -297.01900 (that) -296.98500 (using) -296.98000 (our) -296.99200 (PG) -297.00200 (method) -296.99000 (we) ] TJ 11.95510 TL T* [ (can) -440 (optimize) -440.00900 (any) -439.01900 (of) -439.99300 (the) -439.99800 (metrics\054) -486.99300 (including) -440.00300 (the) -439.99800 (pr) 44.98510 (oposed) ] TJ T* [ (SPIDEr) -443.00600 (metric) -443.00900 (whic) 14.99870 (h) -444 (r) 37.01960 (esults) -442.99700 (in) -443.01200 (ima) 10.01300 (g) 10.00320 (e) -442.99000 (captions) -442.98700 (that) -444.01900 (ar) 36.98650 (e) ] TJ T* [ (str) 44.99250 (ongly) -373.99800 (pr) 36.98520 (eferr) 36.98520 (ed) -374.01500 (by) -374.01500 (human) -373.99300 (r) 14.98400 (ater) 10.01180 (s) -374.01000 (compar) 36.99380 (ed) -374.99600 (to) -374.00400 (captions) ] TJ T* [ (g) 10.00320 (ener) 15.01960 (ated) -308.01400 (by) -309.00700 (the) -308.02000 (same) -308.99700 (model) -308.01100 (b) 20.00160 (ut) -308.01500 (tr) 14.99140 (ained) -308.99000 (to) -308.01500 (optimize) -309.01000 (MLE) ] TJ T* [ (or) -249.98600 (the) -249.99000 (COCO) -250.00500 (metrics\056) ] TJ /R20 11.95520 Tf 36.42890 TL T* [ (1\056) -249.99000 (Intr) 18.01460 (oduction) ] TJ /R22 9.96260 Tf 11.95510 -19.11600 Td [ (Image) -459.99600 (captioning) -460.98300 (is) -459.98800 (the) -459.99400 (task) -461.00800 (of) -460 (describing) -460.01100 (the) -461.01400 (visual) ] TJ -11.95510 -11.95510 Td [ (content) -746.01200 (of) -745.99100 (an) -745.99200 (image) -746.01600 (using) -746.00400 (one) -745.98700 (or) -745.99100 (more) -745.98300 (sentences\056) ] TJ 11.95510 TL T* [ (This) -488.99400 (has) -489.99000 (man) 14.99010 (y) -489.01200 (applications\054) -549.00200 (including) -489.01400 (te) 14.98030 (xt\055based) -490.01900 (image) ] TJ 11.95590 TL T* [ (retrie) 25.01480 (v) 24.98110 (al\054) -211.99800 (accessibility) -202.00400 (for) -202.00700 (blind) -202.01500 (users) -203.01500 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R22 9.96260 Tf 1 0 0 1 200.61800 147.50600 Tm (25) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R22 9.96260 Tf 1 0 0 1 210.58000 147.50600 Tm [ (\135\054) -210.98400 (and) -203.02000 (human\055robot) ] TJ -160.46800 -11.95510 Td [ (interaction) -250.00900 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R22 9.96260 Tf 1 0 0 1 98.53050 135.55100 Tm (6) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R22 9.96260 Tf 1 0 0 1 103.51200 135.55100 Tm (\135\056) Tj -41.44450 -12.14180 Td [ (Most) -438.99500 (methods) -440.00900 (for) -438.98700 (solving) -438.98400 (this) -440.01800 (task) -439.01000 (require) -438.98300 (training) -440.01500 (a) ] TJ -11.95510 -11.95630 Td [ (statistical) -251.99400 (model) -250.99200 (on) -251.98700 (a) -252.00200 (dataset) -251.00500 (of) -251.99600 (\050image\054) -251.98200 (caption\051) -252.01900 (pairs\056) -314.99400 (The) ] TJ 11.95510 TL T* [ (model) -308.98200 (is) -308.99400 (usually) -309.01600 (trained) -308.01500 (to) -308.99500 (maximize) -309.01600 (the) -309 (log) -308.99000 (lik) 10.01790 (elihood) -309.01700 (of) ] TJ ET Q 3.98000 w 0 G 501.12100 904.14800 m 1446.11000 904.14800 l S q 10 0 0 10 0 0 cm BT /R32 5.97760 Tf 1 0 0 1 60.14100 83.81290 Tm [ <03> -0.90058 ] TJ /R22 7.97010 Tf 4.31680 -2.81289 Td [ (W) 79.97960 (ork) -250.01600 (done) -249.99100 (while) -249.98400 (interning) -250.01300 (at) -249.99300 (Google) -250.01700 (Research\056) ] TJ /R22 9.96260 Tf 244.40400 445.02800 Td [ (the) -334.99600 (training) -335.01300 (set\056) -1129.02000 (After) -335.00300 (training\054) -355.99000 (these) -334.99300 (models) -335.01300 (are) -335.01300 (usually) ] TJ 11.95510 TL T* [ (e) 25.01050 (v) 24.98110 (aluated) -407.99800 (by) -408 (computing) -408.00300 (a) -406.99600 (v) 24.98110 (ariety) -408 (of) -408.01000 (dif) 24.98600 (ferent) -408.01500 (metrics) -407.99100 (on) ] TJ T* [ (a) -464.00600 (test) -463.98400 (set\054) -517 (such) -463.98900 (as) -463.99900 (COCO) -464.00600 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R22 9.96260 Tf 1 0 0 1 426.50000 502.11800 Tm (12) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R22 9.96260 Tf 1 0 0 1 436.46300 502.11800 Tm [ (\135\056) -951.99800 (Standard) -464.01600 (metrics) -464.02100 (from) ] TJ -127.60100 -11.95510 Td [ (the) -475.99200 (machine) -475.98200 (translation) -474.98200 (community) -475.98400 (include) -476.01900 (BLEU) -476.01400 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R22 9.96260 Tf 1 0 0 1 529.34400 490.16300 Tm (14) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R22 9.96260 Tf 1 0 0 1 539.30700 490.16300 Tm (\135\054) Tj -230.44500 -11.95470 Td [ (METEOR) -545.98700 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R22 9.96260 Tf 1 0 0 1 358.57500 478.20800 Tm (3) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R22 9.96260 Tf 1 0 0 1 363.55700 478.20800 Tm [ (\135\054) -620.99400 (and) -547.00200 (R) 40.00560 (OUGE) -546.99200 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R22 9.96260 Tf 1 0 0 1 438.06700 478.20800 Tm (11) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R22 9.96260 Tf 1 0 0 1 448.02900 478.20800 Tm [ (\135\056) -1200 (More) -546.99700 (recently) 65.01860 (\054) -620.99000 (the) ] TJ -139.16700 -11.95630 Td [ (CIDEr) -247.00300 (metric) -247.98300 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R22 9.96260 Tf 1 0 0 1 369.12600 466.25200 Tm (20) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R22 9.96260 Tf 1 0 0 1 379.08800 466.25200 Tm [ (\135) -246.98100 (w) 10 (as) -247.01500 (proposed\054) -248.01300 <73706563690263616c6c79> -246.98300 (for) -247.01800 (the) -247.98800 (image) ] TJ -70.22620 -11.95470 Td [ (captioning) -238.00200 (task\056) -306.99000 (W) 79.98660 (e) -238.00400 (shall) -238.01700 (call) -239.01400 (the) -237.99000 (combination) -239.00900 (of) -238 (these) -239.00900 (four) ] TJ 11.95510 TL T* [ (metrics) -346.00100 (\223BCMR\224\054) -346.01600 (for) -345.98400 (short\056) -597.99300 (Unfortunately) 65.00390 (\054) -370.99200 (none) -346.01100 (of) -346.02100 (these) ] TJ T* [ (metrics) -284.99200 (correlate) -285.01600 (strongly) -285.01100 (with) -284.98700 (human) -284.98700 (measures) -286.01600 (of) -285.01100 (caption) ] TJ T* [ (quality) 65.00140 (\056) -489.99400 (In) -309.98500 (f) 9.99588 (act\054) -325.00500 (humans) -309.99700 (score) -309.98700 (lo) 24.98600 (wer) -309.99200 (on) -310.01700 (these) -310.01700 (metrics) -310.00700 (than) ] TJ 11.95590 TL T* [ (the) -368.01000 (methods) -369 (that) -368.00200 (w) 10 (on) -368.00700 (the) -368.00700 (COCO) -369.00200 (2015) -367.99700 (challenge\054) -397.99700 (despite) ] TJ 11.95510 TL T* [ (the) -249.99000 (f) 9.99588 (act) -250 (that) -249.98300 (humans) -250.00700 (are) -250.01000 (still) -250.00200 (much) -250.01700 (better) -249.98800 (at) -249.99300 (this) -250.01200 (task\056) ] TJ 11.95510 -13.29800 Td [ (These) -330.99600 (results) -331.99900 (moti) 25.98830 (v) 24.98110 (ated) -332.01800 (Anderson) -330.98400 (et) -330.99900 (al\056) -553.98100 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R22 9.96260 Tf 1 0 0 1 490.91900 381.22300 Tm (1) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R22 9.96260 Tf 1 0 0 1 495.90000 381.22300 Tm [ (\135) -331.00800 (to) -332.01300 (propose) ] TJ -187.03800 -11.95590 Td [ (the) -582.01300 (SPICE) -582.01800 (metric\056) -1306.98000 (Rather) -582.00600 (than) -582.00600 (directly) -582 (comparing) -583.01100 (a) ] TJ 11.95510 TL T* [ (generated) -295.98700 (sentence) -297.00400 (to) -296.01700 (a) -295.99500 (set) -297 (of) -295.99000 (reference) -296.01400 (sentences) -296.01900 (in) -297 (terms) ] TJ T* [ (of) -521.98900 (syntactic) -522.00600 (agreement\054) -590.98700 (SPICE) -521.98600 <02727374> -522.00600 (parses) -521.98600 (each) -522.00100 (of) -523.01100 (the) ] TJ T* [ (reference) -191.01500 (sentences\054) -203.01000 (and) -192.00200 (then) -191.01200 (uses) -191.01200 (them) -191.98800 (to) -191.01700 (deri) 24.98110 (v) 14.98280 (e) -190.99300 (an) -192.00700 (abstract) ] TJ T* [ (scene) -273.99600 (graph) -273.98600 (representation\056) -381.99000 (The) -273.98900 (generated) -275.01300 (sentence) -273.98900 (is) -274.01800 (then) ] TJ T* [ (also) -439.99000 (parsed\054) -487.99000 (and) -440 (compared) -439.99500 (to) -439.99500 (the) -439.99500 (graph\073) -534.99400 (this) -441 (allo) 24.99090 (ws) -439.99500 (for) ] TJ 11.95590 TL T* [ (a) -450.00800 (comparison) -448.98900 (of) -450 (the) -449.01600 (semantic) -450.01800 (similarity) 64.98430 (\054) -499.01000 (without) -450.00300 (paying) ] TJ 11.95510 TL T* [ (attention) -621 (to) -620.98200 (syntactic) -621.00900 (f) 9.99588 (actors) -620.99000 (\050modulo) -622.01900 (the) -620.98500 (requirement) ] TJ T* [ (that) -510.99800 (the) -511.98600 (generated) -511.01000 (sentence) -511.98800 (be) -511.01300 (parseable\051\056) -1094.01000 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R22 9.96260 Tf 1 0 0 1 501.52900 273.62600 Tm (1) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R22 9.96260 Tf 1 0 0 1 506.51000 273.62600 Tm [ (\135) -511.99800 (sho) 24.99090 (wed) ] TJ -197.64800 -11.95510 Td [ (that) -433.99100 (SPICE) -434.00400 (is) -433.99400 (the) -433.99900 (only) -433.98400 (e) 15.01220 (xisting) -434.00100 (metric) -433.99100 (that) -433.99100 (has) -434.00100 (a) -434.01100 (strong) ] TJ 11.95470 TL T* [ (correlation) -318.00800 (with) -318.00300 (human) -318.98400 (ratings\054) -335.01800 (and) -317.98100 (ranks) -318.01600 (human) -318.98100 (captions) ] TJ T* [ (abo) 14.98280 (v) 14.98280 (e) -250.00200 (algorithms) -249.99000 (submitted) -249.98800 (to) -249.98800 (the) -249.98800 (COCO) -250.00200 (benchmark\056) ] TJ 11.95510 -13.29920 Td [ (Gi) 24.98600 (v) 14.98280 (en) -616.99500 (this) -617.99000 (result\054) -709 (it) -616.98000 (is) -616.98000 (natural) -618 (to) -616.98500 (w) 10.00320 (ant) -616.98500 (to) -618.00500 (directly) ] TJ -11.95510 -11.95470 Td [ (optimize) -462.00600 (SPICE\056) -461.01600 (Ho) 24.98360 (we) 25.01300 (v) 14.98280 (er) 39.98350 (\054) -515.00700 (this) -460.99700 (is) -461.98700 (trick) 14.98770 (y) 65.01370 (\054) -513.98800 (since) -461.99200 (it) -461.00600 (is) -461.98700 (not) ] TJ T* [ (a) -402.01600 (dif) 24.98600 (ferenti) 0.98513 (able) -402.00900 (objecti) 24.99830 (v) 14.98280 (e\056) -765.01000 (In) -400.99200 (this) -401.98700 (paper) 39.99330 (\054) -438.98100 (we) -402.01100 (sho) 24.99090 (w) -400.98700 (that) -401.99700 (it) ] TJ 11.95630 TL T* [ (is) -457.99000 (possible) -456.98500 (to) -457.99000 (use) -457.99500 (polic) 14.99750 (y) -457.01700 (gradient) -457.98200 (\050PG\051) -457.01200 (methods) -458.00700 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R22 9.96260 Tf 1 0 0 1 519.52100 188.59600 Tm (17) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R22 9.96260 Tf 1 0 0 1 529.48400 188.59600 Tm [ (\135) -458.00700 (to) ] TJ -220.62100 -11.95470 Td [ (optimize) -252.98200 (such) -251.98500 (objecti) 24.99830 (v) 14.98280 (es\056) -317.98100 (The) -253.01200 (idea) -251.99200 (of) -253.01700 (using) -252.00700 (PG) -252.99700 (to) -253.00700 (optimize) ] TJ 11.95510 TL T* [ (non) -272.01800 (dif) 24.98600 (ferentiable) -271.98600 (objecti) 25 (v) 14.98280 (es) -271.99400 (for) -271.99400 (image) -272.01800 (captioning) -271.99400 (w) 10.00320 (as) -271.99400 <02727374> ] TJ T* [ (proposed) -379.99100 (in) -380.00300 (the) -380.00800 (MIXER) -379.98800 (paper) -380.01500 (\133) ] TJ ET Q 0 1 0 rg q 10 0 0 10 0 0 cm BT /R22 9.96260 Tf 1 0 0 1 441.78300 152.73100 Tm (15) Tj ET Q 0 g q 10 0 0 10 0 0 cm BT /R22 9.96260 Tf 1 0 0 1 451.74600 152.73100 Tm [ (\135\056) -700.98100 (Ho) 24.98600 (we) 25.01540 (v) 14.98280 (er) 39.98350 (\054) -412.98500 (the) 14.98770 (y) -380.01000 (only) ] TJ -142.88400 -11.95510 Td [ (used) -350.98800 (it) -352.00500 (to) -350.98800 (optimize) -351.00500 (BLEU\0554\054) -377.00600 (which) -352.00500 (is) -350.98600 (not) -350.98100 (correlated) -352.02000 (with) ] TJ 11.95590 TL T* [ (human) -407.00800 (quality) 65.00140 (\056) -781.98500 (Furthermore\054) -446.00900 (when) -406.98100 (we) -406.99100 (tried) -407.02000 (to) -408 (use) -406.98600 (their) ] TJ 11.95510 TL T* [ (method) -226.99200 (to) -227.00900 (optimize) -226.98700 (other) -227.01100 (metrics\054) -232.01500 (such) -227.01100 (as) -227.02100 (CIDEr) -227.98600 (or) -227.02100 (SPICE\054) ] TJ T* [ (we) -444.00700 (got) -443.00700 (poor) -443.99200 (results\054) -492.00900 (since) -443.01400 (their) -443.99500 (method) -442.99500 (\050which) -443.99000 (in) 40.00310 (v) 20.00160 (olv) 14.99750 (es) ] TJ T* [ (an) -343.00200 (intricate) -342.98400 (incremental) -342 (schedule) -342.98700 (which) -342.98700 (mix) 15.00240 (es) -343.00200 (from) -342.99200 (MLE) ] TJ T* [ (training) -325.01400 (to) -324.99200 (full) -324.98700 (PG) -324.98700 (training\051) -325.01900 (is) -324.99000 (not) -324.98500 (v) 14.98280 (ery) -325.99000 (rob) 20.00160 (ust) -324.98500 (and) -325 (require) ] 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 (873) 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 [ 133 751 479 772 ]
/Type /XObject
/Subtype /Form
>>
stream
x+O4PH/VЯ02Qp
0
endstream
endobj
21 0 obj
<<
/Filter /FlateDecode
/Resources 22 0 R
/Length 53008
/Type /XObject
/BBox [ 133 751 479 772 ]
/Subtype /Form
>>
stream
xlK8,8?DKs9mavd
{f-8*2Y@H >חk}y}uf`v)_s}1z#*Gw_gX jow\o'1c|Z^G<
A̺X}ay?IT|y~L.[ {Ȟb\3-3]_'X\竵0{+_۾oY-wj+ B;)Aa=/
1~a(>}m_K' >*Q1:?OPnGsQ_`?mzN
?WBB!lU n'S 1Ipzχ1xF_09,}?e){:̹~Gu7g {0mD`pdS)>;ys`Pµx'O
+혝 oI~!p7̵&m=dC+>5ALV"I1"S*CKha#03L4|@p灟vMRQ==*/m"r>@uqo9ͻ;Vz 'ȳ;IBzWR~ihwa8PlT>]fZ8y52)r빃W{B7G/q)}M]elS(k=5^A] сڸc'w!Ll'M=Ȥ뎓z;#կN>\q6r>|8}}Lr$Mo5:>syls+Ǵ`C?/9{Y||jD u#i^\rJ8a|oY+Rxl<{[؏s迨nYaյ[?/^u;+ ,bhL}wrj:ɴdM8^q|m'aQ6;0k_ar|E8HɃpN('&r`V;Rs볩Nk\&W=DDG#5]~M{VyCSRH9ܜXˏt7߾i9ADگ
USڝ|{(hbIRZXg @6Hɹ0Dq;si\Tv<1?mUŏV7bF
p7Y9sƨ*.uz_LբřsB݃ɹ^'_Yae91f+n%A̶d
Ċc#q
CF=Mt'\&ξ)z;L-4:n}41j:mӶ_)\݇ |T@uAƓxAN6Ql;}Ɲ{~_y"g%]nCi']I~,xa?feoɌr0u?V=phldx.Y;~nZL603> \`ęܓ