File size: 106,389 Bytes
6fa4bc9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 | {
"paper_id": "P13-1006",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T09:33:48.132975Z"
},
"title": "Grounded Language Learning from Video Described with Sentences",
"authors": [
{
"first": "Haonan",
"middle": [],
"last": "Yu",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Purdue University",
"location": {
"addrLine": "465 Northwestern Ave. West Lafayette",
"postCode": "47907-2035",
"region": "IN",
"country": "USA"
}
},
"email": "haonan@haonanyu.com"
},
{
"first": "Jeffrey",
"middle": [
"Mark"
],
"last": "Siskind",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Purdue University",
"location": {
"addrLine": "465 Northwestern Ave. West Lafayette",
"postCode": "47907-2035",
"region": "IN",
"country": "USA"
}
},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "We present a method that learns representations for word meanings from short video clips paired with sentences. Unlike prior work on learning language from symbolic input, our input consists of video of people interacting with multiple complex objects in outdoor environments. Unlike prior computer-vision approaches that learn from videos with verb labels or images with noun labels, our labels are sentences containing nouns, verbs, prepositions, adjectives, and adverbs. The correspondence between words and concepts in the video is learned in an unsupervised fashion, even when the video depicts simultaneous events described by multiple sentences or when different aspects of a single event are described with multiple sentences. The learned word meanings can be subsequently used to automatically generate description of new video.",
"pdf_parse": {
"paper_id": "P13-1006",
"_pdf_hash": "",
"abstract": [
{
"text": "We present a method that learns representations for word meanings from short video clips paired with sentences. Unlike prior work on learning language from symbolic input, our input consists of video of people interacting with multiple complex objects in outdoor environments. Unlike prior computer-vision approaches that learn from videos with verb labels or images with noun labels, our labels are sentences containing nouns, verbs, prepositions, adjectives, and adverbs. The correspondence between words and concepts in the video is learned in an unsupervised fashion, even when the video depicts simultaneous events described by multiple sentences or when different aspects of a single event are described with multiple sentences. The learned word meanings can be subsequently used to automatically generate description of new video.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "People learn language through exposure to a rich perceptual context. Language is grounded by mapping words, phrases, and sentences to meaning representations referring to the world. Siskind (1996) has shown that even with referential uncertainty and noise, a system based on crosssituational learning can robustly acquire a lexicon, mapping words to word-level meanings from sentences paired with sentence-level meanings. However, it did so only for symbolic representations of word-and sentence-level meanings that were not perceptually grounded. An ideal system would not require detailed word-level labelings to acquire word meanings from video but rather could learn language in a largely unsupervised fashion, just as a child does, from video paired with sentences.",
"cite_spans": [
{
"start": 182,
"end": 196,
"text": "Siskind (1996)",
"ref_id": "BIBREF19"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "There has been recent research on grounded language learning. Roy (2002) pairs training sentences with vectors of real-valued features extracted from synthesized images which depict 2D blocks-world scenes, to learn a specific set of features for adjectives, nouns, and adjuncts. Yu and Ballard (2004) paired training images containing multiple objects with spoken name candidates for the objects to find the correspondence between lexical items and visual features. Dominey and Boucher (2005) paired narrated sentences with symbolic representations of their meanings, automatically extracted from video, to learn object names, spatial-relation terms, and event names as a mapping from the grammatical structure of a sentence to the semantic structure of the associated meaning representation. Chen and Mooney (2008) learned the language of sportscasting by determining the mapping between game commentaries and the meaning representations output by a rulebased simulation of the game. Kwiatkowski et al. (2012) present an approach that learns Montaguegrammar representations of word meanings together with a combinatory categorial grammar (CCG) from child-directed sentences paired with first-order formulas that represent their meaning.",
"cite_spans": [
{
"start": 62,
"end": 72,
"text": "Roy (2002)",
"ref_id": "BIBREF16"
},
{
"start": 279,
"end": 300,
"text": "Yu and Ballard (2004)",
"ref_id": "BIBREF25"
},
{
"start": 466,
"end": 492,
"text": "Dominey and Boucher (2005)",
"ref_id": "BIBREF9"
},
{
"start": 793,
"end": 815,
"text": "Chen and Mooney (2008)",
"ref_id": "BIBREF7"
},
{
"start": 985,
"end": 1010,
"text": "Kwiatkowski et al. (2012)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Although most of these methods succeed in learning word meanings from sentential descriptions they do so only for symbolic or simple visual input (often synthesized); they fail to bridge the gap between language and computer vision, i.e., they do not attempt to extract meaning representations from complex visual scenes. On the other hand, there has been research on training object and event models from large corpora of complex images and video in the computer-vision community (Kuznetsova et al., 2012; Sadanand and Corso, 2012; Ordonez et al., 2011; Yao et al., 2010) . However, most such work requires training data that labels individual concepts with individual words (i.e., ob-jects delineated via bounding boxes in images as nouns and events that occur in short video clips as verbs). There is no attempt to model phrasal or sentential meaning, let alone acquire the object or event models from training data labeled with phrasal or sentential annotations. Moreover, such work uses distinct representations for different parts of speech; i.e., object and event recognizers use different representations.",
"cite_spans": [
{
"start": 481,
"end": 506,
"text": "(Kuznetsova et al., 2012;",
"ref_id": "BIBREF13"
},
{
"start": 507,
"end": 532,
"text": "Sadanand and Corso, 2012;",
"ref_id": "BIBREF17"
},
{
"start": 533,
"end": 554,
"text": "Ordonez et al., 2011;",
"ref_id": "BIBREF15"
},
{
"start": 555,
"end": 572,
"text": "Yao et al., 2010)",
"ref_id": "BIBREF24"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In this paper, we present a method that learns representations for word meanings from short video clips paired with sentences. Our work differs from prior work in three ways. First, our input consists of realistic video filmed in an outdoor environment. Second, we learn the entire lexicon, including nouns, verbs, prepositions, adjectives, and adverbs, simultaneously from video described with whole sentences. Third we adopt a uniform representation for the meanings of words in all parts of speech, namely Hidden Markov Models (HMMs) whose states and distributions allow for multiple possible interpretations of a word or a sentence in an ambiguous perceptual context.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "We employ the following representation to ground the meanings of words, phrases, and sentences in video clips. We first run an object detector on each video frame to yield a set of detections, each a subregion of the frame. In principle, the object detector need just detect the objects rather than classify them. In practice, we employ a collection of class-, shape-, pose-, and viewpoint-specific detectors and pool the detections to account for objects whose shape, pose, and viewpoint may vary over time. Our methods can learn to associate a single noun with detections produced by multiple detectors. We then string together detections from individual frames to yield tracks for objects that temporally span the video clip. We associate a feature vector with each frame (detection) of each such track. This feature vector can encode image features (including the identity of the particular detector that produced that detection) that correlate with object class; region color, shape, and size features that correlate with object properties; and motion features, such as linear and angular object position, velocity, and acceleration, that correlate with event properties. We also compute features between pairs of tracks to encode the relative position and motion of the pairs of objects that participate in events that involve two participants. In principle, we can also compute features between tuples of any number of tracks.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Following Yamoto et al. (1992) , Siskind and Morris (1996) , and Starner et al. (1998) , we represent the meaning of an intransitive verb, like jump, as a two-state HMM over the velocity-direction feature, modeling the requirement that the participant move upward then downward. We represent the meaning of a transitive verb, like pick up, as a two-state HMM over both single-object and object-pair features: the agent moving toward the patient while the patient is as rest, followed by the agent moving together with the patient. We extend this general approach to other parts of speech. Nouns, like person, can be represented as one-state HMMs over image features that correlate with the object classes denoted by those nouns. Adjectives, like red, round, and big, can be represented as one-state HMMs over region color, shape, and size features that correlate with object properties denoted by such adjectives. Adverbs, like quickly, can be represented as one-state HMMs over object-velocity features. Intransitive prepositions, like leftward, can be represented as one-state HMMs over velocity-direction features. Static transitive prepositions, like to the left of, can be represented as one-state HMMs over the relative position of a pair of objects. Dynamic transitive prepositions, like towards, can be represented as HMMs over the changing distance between a pair of objects. Note that with this formulation, the representation of a verb, like approach, might be the same as a dynamic transitive preposition, like towards. While it might seem like overkill to represent the meanings of words as one-state-HMMs, in practice, we often instead encode such concepts with multiple states to allow for temporal variation in the associated features due to changing pose and viewpoint as well as deal with noise and occlusion. Moreover, the general framework of modeling word meanings as temporally variant time series via multi-state HMMs allows one to model denominalized verbs, i.e., nouns that denote events, as in The jump was fast.",
"cite_spans": [
{
"start": 10,
"end": 30,
"text": "Yamoto et al. (1992)",
"ref_id": "BIBREF23"
},
{
"start": 33,
"end": 58,
"text": "Siskind and Morris (1996)",
"ref_id": "BIBREF18"
},
{
"start": 61,
"end": 86,
"text": "and Starner et al. (1998)",
"ref_id": "BIBREF20"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Our HMMs are parameterized with varying arity.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Some, like jump(\u03b1), person(\u03b1), red(\u03b1), round(\u03b1), big(\u03b1), quickly(\u03b1), and leftward(\u03b1) have one argument, while others, like pick-up(\u03b1, \u03b2), to-the-left-of(\u03b1, \u03b2), and towards(\u03b1, \u03b2), have two arguments (In principle, any arity can be supported.). HMMs are instantiated by mapping their arguments to tracks. This involves computing the associated feature vector for that HMM over the detections in the tracks chosen to fill its arguments. This is done with a two-step process to support compositional semantics. The meaning of a multi-word phrase or sentence is represented as a joint likelihood of the HMMs for the words in that phrase or sentence. Compositionality is handled by linking or coindexing the arguments of the conjoined HMMs. Thus a sentence like The person to the left of the backpack approached the trashcan would be represented as a conjunction of person(p 0 ), to-the-left-of(p 0 , p 1 ), backback(p 1 ), approached(p 0 , p 2 ), and trash-can(p 2 ) over the three participants p 0 , p 1 , and p 2 . This whole sentence is then grounded in a particular video by mapping these participants to particular tracks and instantiating the associated HMMs over those tracks, by computing the feature vectors for each HMM from the tracks chosen to fill its arguments.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Our algorithm makes six assumptions. First, we assume that we know the part of speech C m associated with each lexical entry m, along with the part-of-speech dependent number of states I c in the HMMs used to represent word meanings in that part of speech, the part-of-speech dependent number of features N c in the feature vectors used by HMMs to represent word meanings in that part of speech, and the part-of-speech dependent feature-vector computation \u03a6 c used to compute the features used by HMMs to represent word meanings in that part of speech. Second, we pair individual sentences each with a short video clip that depicts that sentence. The algorithm is not able to determine the alignment between multiple sentences and longer video segments. Note that there is no requirement that the video depict only that sentence. Other objects may be present and other events may occur. In fact, nothing precludes a training corpus with multiple copies of the same video, each paired with a different sentence describing a different aspect of that video. Moreover, our algorithm potentially can handle a small amount of noise, where a video clip is paired with an incorrect sentence that the video does not depict. Third, we assume that we already have (pre-trained) low-level object detectors capable of detecting instances of our target event participants in individual frames of the video. We allow such detections to be unreliable; our method can handle a moderate amount of false positives and false negatives. We do not need to know the mapping from these object-detection classes to words; our algorithm determines that. Fourth, we assume that we know the arity of each word in the corpus, i.e., the number of arguments that that word takes. For example, we assume that we know that the word person(\u03b1) takes one argument and the word approached(\u03b1, \u03b2) takes two arguments. Fifth, we assume that we know the total number of distinct participants that collectively fill all of the arguments for all of the words in each training sentence. For example, for the sentence The person to the left of the backpack approached the trash-can, we assume that we know that there are three distinct objects that participate in the event denoted. Sixth, we assume that we know the argument-to-participant mapping for each training sentence. Thus, for example, for the above sentence we would know",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "person(p 0 ), to-the-left-of(p 0 , p 1 ), backback(p 1 ), approached(p 0 , p 2 )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": ", and trash-can(p 2 ). The latter two items can be determined by parsing the sentence, which is what we do. One can imagine learning the ability to automatically perform the latter two items, and even the fourth item above, by learning the grammar and the part of speech of each word, such as done by Kwiatkowski et al. (2012) . We leave such for future work. Figure 1 illustrates a single frame from a potential training sample provided as input to our learner. It consists of a video clip paired with a sentence, where the arguments of the words in the sentence are mapped to participants. From a sequence of such training samples, our learner determines the objects tracks and the mapping from participants to those tracks, together with the meanings of the words.",
"cite_spans": [
{
"start": 301,
"end": 326,
"text": "Kwiatkowski et al. (2012)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [
{
"start": 360,
"end": 368,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The remainder of the paper is organized as follows. Section 2 generally describes our problem of lexical acquisition from video. Section 3 introduces our work on the sentence tracker, a method for jointly tracking the motion of multiple objects in a video that participate in a sententiallyspecified event. Section 4 elaborates on the details of our problem formulation in the context of this sentence tracker. Section 5 describes how to generalize and extend the sentence tracker so that it can be used to support lexical acquisition. We demonstrate this lexical acquisition algorithm on a small example in Section 6. Finally, we conclude with a discussion in Section 7. Figure 1: An illustration of our problem. Each word in the sentence has one or more arguments (\u03b1 and possibly \u03b2), each argument of each word is assigned to a participant (p 0 , . . . , p 3 ) in the event described by the sentence, and each participant can be assigned to any object track in the video. This figure shows a possible (but erroneous) interpretation of the sentence where the mapping is:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "p 0 \u2192 Track 3, p 1 \u2192 Track 0, p 2 \u2192 Track 1,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "and p 3 \u2192 Track 2, which might (incorrectly) lead the learner to conclude that the word person maps to the backpack, the word backpack maps to the chair, the word trash-can maps to the trash-can, and the word chair maps to the person.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Throughout this paper, lowercase letters are used for variables or hidden quantities while uppercase ones are used for constants or observed quantities. We are given a lexicon {1, . . . , M }, letting m denote a lexical entry. We are given a sequence D = (D 1 , . . . , D R ) of video clips D r , each paired with a sentence S r from a sequence S = (S 1 , . . . , S R ) of sentences. We refer to D r paired with S r as a training sample. Each sentence S r is a sequence (S r,1 , . . . , S r,Lr ) of words S r,l , each an entry from the lexicon. A given entry may potentially appear in multiple sentences and even multiple times in a given sentence. For example, the third word in the first sentence might be the same entry as the second word in the fourth sentence, in which case S 1,3 = S 4,2 . This is what allows cross-situational learning in our algorithm.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "General Problem Formulation",
"sec_num": "2"
},
{
"text": "Let us assume, for a moment, that we can process each video clip D r to yield a sequence (\u03c4 r,1 , . . . , \u03c4 r,Ur ) of object tracks \u03c4 r,u . Let us also assume that D r is paired with a sen-tence S r = The person approached the chair, specified to have two participants, p r,0 and p r,1 , with the mapping person(p r,0 ), chair(p r,1 ), and approached(p r,0 , p r,1 ). Let us further assume, for a moment, that we are given a mapping from participants to object tracks, say p r,0 \u2192 \u03c4 r,39 and p r,1 \u2192 \u03c4 r,51 . This would allow us to instantiate the HMMs with object tracks for a given video clip: person(\u03c4 r,39 ), chair(\u03c4 r,51 ), and approached(\u03c4 r,39 , \u03c4 r,51 ). Let us further assume that we can score each such instantiated HMM and aggregate the scores for all of the words in a sentence to yield a sentence score and further aggregate the scores for all of the sentences in the corpus to yield a corpus score. However, we don't know the parameters of the HMMs. These constitute the unknown meanings of the words in our corpus which we wish to learn. The problem is to simultaneously determine (a) those parameters along with (b) the object tracks and (c) the mapping from participants to object tracks. We do this by finding (a)-(c) that maximizes the corpus score.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "General Problem Formulation",
"sec_num": "2"
},
{
"text": "3 The Sentence Tracker Barbu et al. (2012a) presented a method that first determines object tracks from a single video clip and then uses these fixed tracks with HMMs to recognize actions corresponding to verbs and construct sentential descriptions with templates. Later Barbu et al. (2012b) addressed the problem of solving (b) and (c), for a single object track constrained by a single intransitive verb, without solving (a), in the context of a single video clip. Our group has generalized this work to yield an algorithm called the sentence tracker which operates by way of a factorial HMM framework. We introduce that here as the foundation of our extension.",
"cite_spans": [
{
"start": 23,
"end": 43,
"text": "Barbu et al. (2012a)",
"ref_id": "BIBREF0"
},
{
"start": 271,
"end": 291,
"text": "Barbu et al. (2012b)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "General Problem Formulation",
"sec_num": "2"
},
{
"text": "Each video clip D r contains T r frames. We run an object detector on each frame to yield a set D t r of detections. Since our object detector is unreliable, we bias it to have high recall but low precision, yielding multiple detections in each frame. We form an object track by selecting a single detection for that track for each frame. For a moment, let us consider a single video clip with length T , with detections D t in frame t. Further, let us assume that we seek a single object track in that video clip. Let j t denote the index of the detection from D t in frame t that is selected to form the track. The object detector scores each detection. Let F (D t , j t ) denote that score. More-over, we wish the track to be temporally coherent; we want the objects in a track to move smoothly over time and not jump around the field of view. Let G(D t\u22121 , j t\u22121 , D t , j t ) denote some measure of coherence between two detections in adjacent frames. (One possible such measure is consistency of the displacement of D t relative to D t\u22121 with the velocity of D t\u22121 computed from the image by optical flow.) One can select the detections to yield a track that maximizes both the aggregate detection score and the aggregate temporal coherence score.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "General Problem Formulation",
"sec_num": "2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "max j 1 ,...,j T \uf8eb \uf8ec \uf8ec \uf8ec \uf8ec \uf8ed T t=1 F (D t , j t ) + T t=2 G(D t\u22121 , j t\u22121 , D t , j t ) \uf8f6 \uf8f7 \uf8f7 \uf8f7 \uf8f7 \uf8f8",
"eq_num": "(1)"
}
],
"section": "General Problem Formulation",
"sec_num": "2"
},
{
"text": "This can be determined with the Viterbi (1967) algorithm and is known as detection-based tracking (Viterbi, 1971) . Recall that we model the meaning of an intransitive verb as an HMM over a time series of features extracted for its participant in each frame. Let \u03bb denote the parameters of this HMM, (q 1 , . . . , q T ) denote the sequence of states q t that leads to an observed track, B(D t , j t , q t , \u03bb) denote the conditional log probability of observing the feature vector associated with the detection selected by j t among the detections D t in frame t, given that the HMM is in state q t , and A(q t\u22121 , q t , \u03bb) denote the log transition probability of the HMM. For a given track (j 1 , . . . , j T ), the state sequence that yields the maximal likelihood is given by:",
"cite_spans": [
{
"start": 32,
"end": 46,
"text": "Viterbi (1967)",
"ref_id": "BIBREF21"
},
{
"start": 98,
"end": 113,
"text": "(Viterbi, 1971)",
"ref_id": "BIBREF22"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "General Problem Formulation",
"sec_num": "2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "max q 1 ,...,q T \uf8eb \uf8ec \uf8ec \uf8ec \uf8ec \uf8ed T t=1 B(D t , j t , q t , \u03bb) + T t=2 A(q t\u22121 , q t , \u03bb) \uf8f6 \uf8f7 \uf8f7 \uf8f7 \uf8f7 \uf8f8",
"eq_num": "(2)"
}
],
"section": "General Problem Formulation",
"sec_num": "2"
},
{
"text": "which can also be found by the Viterbi algorithm. A given video clip may depict multiple objects, each moving along its own trajectory. There may be both a person jumping and a ball rolling. How are we to select one track over the other? The key insight of the sentence tracker is to bias the selection of a track so that it matches an HMM. This is done by combining the cost function of Eq. 1 with the cost function of Eq. 2 to yield Eq. 3, which can also be determined using the Viterbi algorithm. This is done by forming the cross product of the two lattices. This jointly selects the optimal detections to form the track, together with the optimal state sequence, and scores that combination.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "General Problem Formulation",
"sec_num": "2"
},
{
"text": "max j 1 ,...,j T q 1 ,...,q T \uf8eb \uf8ec \uf8ec \uf8ec \uf8ec \uf8ec \uf8ed T t=1 F (D t , j t ) + B(D t , j t , q t , \u03bb) + T t=2 G(D t\u22121 , j t\u22121 , D t , j t ) + A(q t\u22121 , q t , \u03bb) \uf8f6 \uf8f7 \uf8f7 \uf8f7 \uf8f7 \uf8f7 \uf8f8 (3)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "General Problem Formulation",
"sec_num": "2"
},
{
"text": "While we formulated the above around a single track and a word that contains a single participant, it is straightforward to extend this so that it supports multiple tracks and words of higher arity by forming a larger cross product. When doing so, we generalize j t to denote a sequence of detections from D t , one for each of the tracks. We further need to generalize F so that it computes the joint score of a sequence of detections, one for each track, G so that it computes the joint measure of coherence between a sequence of pairs of detections in two adjacent frames, and B so that it computes the joint conditional log probability of observing the feature vectors associated with the sequence of detections selected by j t . When doing this, note that Eqs. 1 and 3 maximize over j 1 , . . . , j T which denotes T sequences of detection indices, rather than T individual indices.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "General Problem Formulation",
"sec_num": "2"
},
{
"text": "It is further straightforward to extend the above to support a sequence (S 1 , . . . , S L ) of words S l denoting a sentence, each of which applies to different subsets of the multiple tracks, again by forming a larger cross product. When doing so, we generalize q t to denote a sequence (q t 1 , . . . , q t L ) of states q t l , one for each word l in the sentence, and use q l to denote the sequence (q 1 l , . . . , q T l ) and q to denote the sequence (q 1 , . . . , q L ). We further need to generalize B so that it computes the joint conditional log probability of observing the feature vectors for the detections in the tracks that are assigned to the arguments of the HMM for each word in the sentence and A so that it computes the joint log transition probability for the HMMs for all words in the sentence. This allows selection of an optimal sequence of tracks that yields the highest score for the sentential meaning of a sequence of words. Modeling the meaning of a sentence through a sequence of words whose meanings are modeled by HMMs, defines a factorial HMM for that sentence, since the overall Markov process for that sentence can be factored into inde-pendent component processes (Brand et al., 1997; Zhong and Ghosh, 2001 ) for the individual words. In this view, q denotes the state sequence for the combined factorial HMM and q l denotes the factor of that state sequence for word l. The remainder of this paper wraps this sentence tracker in Baum Welch (Baum et al., 1970; Baum, 1972) .",
"cite_spans": [
{
"start": 1202,
"end": 1222,
"text": "(Brand et al., 1997;",
"ref_id": "BIBREF6"
},
{
"start": 1223,
"end": 1244,
"text": "Zhong and Ghosh, 2001",
"ref_id": "BIBREF26"
},
{
"start": 1479,
"end": 1498,
"text": "(Baum et al., 1970;",
"ref_id": "BIBREF3"
},
{
"start": 1499,
"end": 1510,
"text": "Baum, 1972)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "General Problem Formulation",
"sec_num": "2"
},
{
"text": "We adapt the sentence tracker to training a corpus of R video clips, each paired with a sentence. Thus we augment our notation, generalizing j t to j t r and q t l to q t r,l . Below, we use j r to denote (j 1 r , . . . , j Tr r ), j to denote (j 1 , . . . , j R ), q r,l to denote (q 1 r,l , . . . , q Tr r,l ), q r to denote (q r,1 , . . . , q r,Lr ), and q to denote (q 1 , . . . , q R ).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Detailed Problem Formulation",
"sec_num": "4"
},
{
"text": "We use discrete features, namely natural numbers, in our feature vectors, quantized by a binning process. We assume the part of speech of entry m is known as C m . The length of the feature vector may vary across parts of speech. Let N c denote the length of the feature vector for part of speech c, x r,l denote the time-series (x 1 r,l , . . . , x Tr r,l ) of feature vectors x t r,l , associated with S r,l (which recall is some entry m), and x r denote the sequence (x r,1 , . . . , x r,Lr ). We assume that we are given a function \u03a6 c (D t r , j t r ) that computes the feature vector x t r,l for the word S r,l whose part of speech is C S r,l = c. Note that we allow \u03a6 to be dependent on c allowing different features to be computed for different parts of speech, since we can determine m and thus C m from S r,l . We choose to have N c and \u03a6 c depend on the part of speech c and not on the entry m since doing so would be tantamount to encoding the to-be-learned word meaning in the provided feature-vector computation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Detailed Problem Formulation",
"sec_num": "4"
},
{
"text": "The goal of training is to find a sequence \u03bb = (\u03bb 1 , . . . , \u03bb M ) of parameters \u03bb m that best explains the R training samples. The parameters \u03bb m constitute the meaning of the entry m in the lexicon. Collectively, these are the initial state probabilities a m 0,k , for 1 \u2264 k \u2264 I Cm , the transition probabilities a m i,k , for 1 \u2264 i, k \u2264 I Cm , and the output probabilities b m i,n (x), for 1 \u2264 i \u2264 I Cm and 1 \u2264 n \u2264 N Cm , where I Cm denotes the number of states in the HMM for entry m. Like before, we could have a distinct I m for each entry m but instead have I Cm depend only on the part of speech of entry m, and assume that we know the fixed I for each part of speech. In our case, b m i,n is a discrete distribution because the features are binned.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Detailed Problem Formulation",
"sec_num": "4"
},
{
"text": "Instantiating the above approach requires a definition for what it means to best explain the R training samples. Towards this end, we define the score of a video clip D r paired with sentence S r given the parameter set \u03bb to characterize how well this training sample is explained. While the cost function in Eq. 3 may qualify as a score, it is easier to fit a likelihood calculation into the Baum-Welch framework than a MAP estimate. Thus we replace the max in Eq. 3 with a and redefine our scoring function as follows:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The Learning Algorithm",
"sec_num": "5"
},
{
"text": "L(D r ; S r , \u03bb) = jr P (j r |D r )P (x r |S r , \u03bb) (4)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The Learning Algorithm",
"sec_num": "5"
},
{
"text": "The score in Eq. 4 can be interpreted as an expectation of the HMM likelihood over all possible mappings from participants to all possible tracks. By definition, P (j r |D r ) = V (Dr,jr) j r V (Dr,j r ) , where the numerator is the score of a particular track sequence j r while the denominator sums the scores over all possible track sequences. The log of the numerator V (D r , j r ) is simply Eq. 1 without the max. The log of the denominator can be computed efficiently by the forward algorithm (Baum and Petrie, 1966) . The likelihood for a factorial HMM can be computed as:",
"cite_spans": [
{
"start": 178,
"end": 187,
"text": "V (Dr,jr)",
"ref_id": null
},
{
"start": 500,
"end": 523,
"text": "(Baum and Petrie, 1966)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "The Learning Algorithm",
"sec_num": "5"
},
{
"text": "P (x r |S r , \u03bb) = qr l P (x r,l , q r,l |S r,l , \u03bb) (5)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The Learning Algorithm",
"sec_num": "5"
},
{
"text": "i.e., summing the likelihoods for all possible state sequences. Each summand is simply the joint likelihood for all the words in the sentence conditioned on a state sequence q r . For HMMs we have",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The Learning Algorithm",
"sec_num": "5"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "P (x r,l , q r,l |S r,l , \u03bb) = t a S r,l q t\u22121 r,l ,q t r,l n b S r,l q t r,l ,n (x t r,l,n )",
"eq_num": "(6)"
}
],
"section": "The Learning Algorithm",
"sec_num": "5"
},
{
"text": "Finally, for a training corpus of R samples, we seek to maximize the joint score:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The Learning Algorithm",
"sec_num": "5"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "L(D; S, \u03bb) = r L(D r ; S r , \u03bb)",
"eq_num": "(7)"
}
],
"section": "The Learning Algorithm",
"sec_num": "5"
},
{
"text": "A local maximum can be found by employing the Baum-Welch algorithm (Baum et al., 1970; Baum, 1972) . By constructing an auxiliary function (Bilmes, 1997) , one can derive the reestimation formulas in Eq. 8, where x t r,l,n = h denotes the selection of all possible j t r such that the nth",
"cite_spans": [
{
"start": 67,
"end": 86,
"text": "(Baum et al., 1970;",
"ref_id": "BIBREF3"
},
{
"start": 87,
"end": 98,
"text": "Baum, 1972)",
"ref_id": "BIBREF4"
},
{
"start": 139,
"end": 153,
"text": "(Bilmes, 1997)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "The Learning Algorithm",
"sec_num": "5"
},
{
"text": "a m i,k = \u03b8 m i R r=1 Lr l=1 s.t.S r,l =m Tr t=1 L(q t\u22121 r,l = i, q t r,l = k, D r ; S r , \u03bb ) L(D r ; S r , \u03bb ) \u03be(r,l,i,k,t) b m i,n (h) = \u03c8 m i,n R r=1 Lr l=1 s.t.S r,l =m Tr t=1 L(q t r,l = i, x t r,l,n = h, D r ; S r , \u03bb ) L(D r ; S r , \u03bb ) \u03b3(r,l,n,i,h,t) (8) feature computed by \u03a6 Cm (D t r , j t r )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The Learning Algorithm",
"sec_num": "5"
},
{
"text": "is h. The coefficients \u03b8 m i and \u03c8 m i,n are for normalization. The reestimation formulas involve occurrence counting. However, since we use a factorial HMM that involves a cross-product lattice and use a scoring function derived from Eq. 3 that incorporates both tracking (Eq. 1) and word models (Eq. 2), we need to count the frequency of transitions in the whole cross-product lattice. As an example of such cross-product occurrence counting, when counting the transitions from state i to k for the lth word from frame t \u2212 1 to t, i.e., \u03be(r, l, i, k, t), we need to count all the possible paths through the adjacent factorial states",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The Learning Algorithm",
"sec_num": "5"
},
{
"text": "(j t\u22121 r , q t\u22121 r,1 , . . . , q t\u22121 r,Lr ) and (j t r , q t r,1 , . . . , q t r,Lr ) such that q t\u22121 r,l = i and q t r,l = k.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The Learning Algorithm",
"sec_num": "5"
},
{
"text": "Similarly, when counting the frequency of being at state i while observing h as the nth feature in frame t for the lth word of entry m, i.e., \u03b3(r, l, n, i, h, t), we need to count all the possible paths through the factorial state (j t r , q t r,1 , . . . , q t r,Lr ) such that q t r,l = i and the nth feature computed by \u03a6 Cm (D t r , j t r ) is h. The reestimation of a single component HMM can depend on the previous estimate for other component HMMs. This dependence happens because of the argument-to-participant mapping which coindexes arguments of different component HMMs to the same track. It is precisely this dependence that leads to cross-situational learning of two kinds: both inter-sentential and intra-sentential. Acquisition of a word meaning is driven across sentences by entries that appear in more than one training sample and within sentences by the requirement that the meanings of all of the individual words in a sentence be consistent with the collective sentential meaning.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The Learning Algorithm",
"sec_num": "5"
},
{
"text": "We filmed 61 video clips (each 3-5 seconds at 640\u00d7480 resolution and 40 fps) that depict a variety of different compound events. Each clip depicts multiple simultaneous events between some Table 1 : The grammar used for our annotation and generation. Our lexicon contains 1 determiner, 4 nouns, 2 spatial relation prepositions, 4 verbs, 2 adverbs, and 2 motion prepositions for a total of 15 lexical entries over 6 parts of speech. subset of four objects: a person, a backpack, a chair, and a trash-can. These clips were filmed in three different outdoor environments which we use for cross validation. We manually annotated each video with several sentences that describe what occurs in that video. The sentences were constrained to conform to the grammar in Table 1 . Our corpus of 159 training samples pairs some videos with more than one sentence and some sentences with more than one video, with an average of 2.6 sentences per video 1 .",
"cite_spans": [],
"ref_spans": [
{
"start": 189,
"end": 196,
"text": "Table 1",
"ref_id": null
},
{
"start": 760,
"end": 767,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Experiment",
"sec_num": "6"
},
{
"text": "S \u2192 NP VP NP \u2192 D N [PP] D \u2192 the N \u2192 person | backpack | trash-can | chair PP \u2192 P NP P \u2192 to the left of | to the right of VP \u2192 V NP [ADV] [PPM] V \u2192 picked up | put down | carried | approached ADV \u2192 quickly | slowly PPM \u2192 PM NP PM \u2192 towards | away from",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiment",
"sec_num": "6"
},
{
"text": "We model and learn the semantics of all words except determiners. Table 2 specifies the arity, the state number I c , and the features computed by \u03a6 c for the semantic models for words of each part of speech c. While we specify a different subset of features for each part of speech, we presume that, in principle, with enough training data, we could include all features in all parts of speech and automatically learn which ones are noninformative and lead to uniform distributions.",
"cite_spans": [],
"ref_spans": [
{
"start": 66,
"end": 73,
"text": "Table 2",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Experiment",
"sec_num": "6"
},
{
"text": "We use an off-the-shelf object detector (Felzenszwalb et al., 2010a; Felzenszwalb et al., 2010b) which outputs detections in the form of scored axis-aligned rectangles. We trained four object detectors, one for each of the four object classes in our corpus: person, backpack, chair, and trashcan. For each frame, we pick the two highestscoring detections produced by each object detector and pool the results yielding eight detections per frame. Having a larger pool of detections per frame can better compensate for false negatives in the object detection and potentially yield smoother tracks but it increases the size of the lattice and the concomitant running time and does not lead to appreciably better performance on our corpus. We compute continuous features, such as velocity, distance, size ratio, and x-position solely from the detection rectangles and quantize the features into bins as follows: velocity To reduce noise, we compute the velocity of a participant by averaging the optical flow in the detection rectangle. The velocity magnitude is quantized into 5 levels: absolutely stationary, stationary, moving, fast moving, and quickly. The velocity orientation is quantized into 4 directions: left, up, right, and down. distance We compute the Euclidean distance between the detection centers of two participants, which is quantized into 3 levels: near, normal, and far away. size ratio We compute the ratio of detection area of the first participant to the detection area of the second participant, quantized into 2 possibilities: larger/smaller than. x-position We compute the difference between the x-coordinates of the participants, quantized into 2 possibilities: to the left/right of. The binning process was determined by a preprocessing step that clustered a subset of the training data. We also incorporate the index of the detector that produced the detection as a feature. The par-ticular features computed for each part of speech are given in Table 2 . Note that while we use English phrases, like to the left of, to refer to particular bins of particular features, and we have object detectors which we train on samples of a particular object class such as backpack, such phrases are only mnemonic of the clustering and object-detector training process. We do not have a fixed correspondence between the lexical entries and any particular feature value. Moreover, that correspondence need not be oneto-one: a given lexical entry may correspond to a (time variant) constellation of feature values and any given feature value may participate in the meaning of multiple lexical entries.",
"cite_spans": [
{
"start": 40,
"end": 68,
"text": "(Felzenszwalb et al., 2010a;",
"ref_id": "BIBREF10"
},
{
"start": 69,
"end": 96,
"text": "Felzenszwalb et al., 2010b)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [
{
"start": 1972,
"end": 1979,
"text": "Table 2",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Experiment",
"sec_num": "6"
},
{
"text": "We perform a three-fold cross validation, taking the test data for each fold to be the videos filmed in a given outdoor environment and the training data for that fold to be all training samples that contain other videos. For testing, we hand selected 24 sentences generated by the grammar in Table 1 , where each sentence is true for at least one test video. Half of these sentences (designated NV) contain only nouns and verbs while the other half (designated ALL) contain other parts of speech. The latter are longer and more complicated than the former. We score each testing video paired with every sentence in both NV and ALL. To evaluate our results, we manually annotated the correctness of each such pair.",
"cite_spans": [],
"ref_spans": [
{
"start": 293,
"end": 300,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Experiment",
"sec_num": "6"
},
{
"text": "Video-sentence pairs could be scored with Eq. 4. However, the score depends on the sentence length, the collective numbers of states and features in the HMMs for words in that sentence, and the length of the video clip. To render the scores comparable across such variation we incorporate a sentence prior to the per-frame score:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiment",
"sec_num": "6"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "L(D r , S r ; \u03bb) = [L(D r ; S r , \u03bb)] 1 Tr \u03c0(S r ) (9) where \u03c0(S r ) = exp Lr l=1 \uf8eb \uf8ec \uf8ec \uf8ed E(I C S r,l ) + N C S r,l n=1 E(Z C S r,l ,n ) \uf8f6 \uf8f7 \uf8f7 \uf8f8",
"eq_num": "(10)"
}
],
"section": "Experiment",
"sec_num": "6"
},
{
"text": "In the above, Z C S r,l ,n is the number of bins for the nth feature of S r,l of part of speech C S r,l and",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiment",
"sec_num": "6"
},
{
"text": "E(Y ) = \u2212 Y y=1 1",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiment",
"sec_num": "6"
},
{
"text": "Y log 1 Y = log Y is the entropy of a uniform distribution over Y bins. This prior prefers longer sentences which describe more information in the video. The scores are thresholded to decide hits, which together with the manual annotations, can generate TP, TN, FP, and FN counts. We select the threshold that leads to the maximal F1 score on the training set, use this threshold to compute F1 scores on the test set in each fold, and average F1 scores across the folds.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiment",
"sec_num": "6"
},
{
"text": "The F1 scores are listed in the column labeled Our in Table 3 . For comparison, we also report F1 scores for three baselines: Chance, Blind, and Hand. The Chance baseline randomly classifies a video-sentence pair as a hit with probability 0.5. The Blind baseline determines hits by potentially looking at the sentence but never looking at the video. We can find an upper bound on the F1 score that any blind method could have on each of our test sets by solving a 0-1 fractional programming problem (Dinkelbach, 1967 ) (see Appendix A for details). The Hand baseline determines hits with hand-coded HMMs, carefully designed to yield what we believe is near-optimal performance. As can be seen from Table 3 , our trained models perform substantially better than the Chance and Blind baselines and approach the performance of the ideal Hand baseline. One can further see from the ROC curves in Figure 2 , comparing the trained and hand-written models on both NV and ALL, that the trained models are close to optimal. Note that performance on ALL exceeds that on NV with the trained models. This is because longer sentences with varied parts of speech incorporate more information into the scoring process.",
"cite_spans": [
{
"start": 499,
"end": 516,
"text": "(Dinkelbach, 1967",
"ref_id": "BIBREF8"
}
],
"ref_spans": [
{
"start": 54,
"end": 61,
"text": "Table 3",
"ref_id": "TABREF3"
},
{
"start": 698,
"end": 705,
"text": "Table 3",
"ref_id": "TABREF3"
},
{
"start": 892,
"end": 900,
"text": "Figure 2",
"ref_id": "FIGREF2"
}
],
"eq_spans": [],
"section": "Experiment",
"sec_num": "6"
},
{
"text": "We presented a method that learns word meanings from video paired with sentences. Unlike prior work, our method deals with realistic video scenes labeled with whole sentences, not individual words labeling hand delineated objects or events. The experiment shows that it can correctly learn the meaning representations in terms of HMM parameters for our lexical entries, from highly ambiguous training data. Our maximumlikelihood method makes use of only positive sentential labels. As such, it might require more training data for convergence than a method that also makes use of negative training sentences that are not true of a given video. Such can be handled with discriminative training, a topic we plan to address in the future. We believe that this will allow learning larger lexicons from more complex video without excessive amounts of training data. 0060. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either express or implied, of the Army Research Laboratory or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes, notwithstanding any copyright notation herein.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "7"
},
{
"text": "A Blind algorithm makes identical decisions on the same sentence paired with different video clips. An optimal algorithm will try to find a decision s i for each test sentence i that maximizes the F1 score. Suppose, the ground-truth yields FP i false positives and TP i true positives on the test set when s i = 1. Also suppose that setting s i = 0 yields FN i false negatives. Then the F1 score is",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "A An Upper Bound on the F1 Score of any Blind Method",
"sec_num": null
},
{
"text": "F 1 = 1 1 + i s i FP i + (1 \u2212 s i )FN i i 2s i TP i \u2206",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "A An Upper Bound on the F1 Score of any Blind Method",
"sec_num": null
},
{
"text": "Thus we want to minimize the term \u2206. This is an instance of a 0-1 fractional programming problem which can be solved by binary search or Dinkelbach's algorithm (Dinkelbach, 1967) .",
"cite_spans": [
{
"start": 160,
"end": 178,
"text": "(Dinkelbach, 1967)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "A An Upper Bound on the F1 Score of any Blind Method",
"sec_num": null
},
{
"text": "Our code, videos, and sentential annotations are available at http://haonanyu.com/research/ acl2013/.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "This research was sponsored by the Army Research Laboratory and was accomplished under Cooperative Agreement Number W911NF-10-2-",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgments",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Video in sentences out",
"authors": [
{
"first": "A",
"middle": [],
"last": "Barbu",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Bridge",
"suffix": ""
},
{
"first": "Z",
"middle": [],
"last": "Burchill",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Coroian",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Dickinson",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Fidler",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Michaux",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Mussman",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Siddharth",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Salvi",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Schmidt",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Shangguan",
"suffix": ""
},
{
"first": "J",
"middle": [
"M"
],
"last": "Siskind",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Waggoner",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Wei",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Yin",
"suffix": ""
},
{
"first": "Z",
"middle": [],
"last": "Zhang",
"suffix": ""
}
],
"year": 2012,
"venue": "Proceedings of the Twenty-Eighth Conference on Uncertainty in Artificial Intelligence",
"volume": "",
"issue": "",
"pages": "102--112",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "A. Barbu, A. Bridge, Z. Burchill, D. Coroian, S. Dick- inson, S. Fidler, A. Michaux, S. Mussman, N. Sid- dharth, D. Salvi, L. Schmidt, J. Shangguan, J. M. Siskind, J. Waggoner, S. Wang, J. Wei, Y. Yin, and Z. Zhang. 2012a. Video in sentences out. In Pro- ceedings of the Twenty-Eighth Conference on Un- certainty in Artificial Intelligence, pages 102-112.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Simultaneous object detection, tracking, and event recognition",
"authors": [
{
"first": "A",
"middle": [],
"last": "Barbu",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Siddharth",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Michaux",
"suffix": ""
},
{
"first": "J",
"middle": [
"M"
],
"last": "Siskind",
"suffix": ""
}
],
"year": 2012,
"venue": "Advances in Cognitive Systems",
"volume": "2",
"issue": "",
"pages": "203--220",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "A. Barbu, N. Siddharth, A. Michaux, and J. M. Siskind. 2012b. Simultaneous object detection, tracking, and event recognition. Advances in Cognitive Systems, 2:203-220, December.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Statistical inference for probabilistic functions of finite state Markov chains",
"authors": [
{
"first": "L",
"middle": [
"E"
],
"last": "Baum",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Petrie",
"suffix": ""
}
],
"year": 1966,
"venue": "The Annals of Mathematical Statistics",
"volume": "37",
"issue": "",
"pages": "1554--1563",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "L. E. Baum and T. Petrie. 1966. Statistical inference for probabilistic functions of finite state Markov chains. The Annals of Mathematical Statistics, 37:1554-1563.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "A maximization technique occuring in the statistical analysis of probabilistic functions of Markov chains",
"authors": [
{
"first": "L",
"middle": [
"E"
],
"last": "Baum",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Petrie",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Soules",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Weiss",
"suffix": ""
}
],
"year": 1970,
"venue": "The Annals of Mathematical Statistics",
"volume": "41",
"issue": "1",
"pages": "164--171",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "L. E. Baum, T. Petrie, G. Soules, and N. Weiss. 1970. A maximization technique occuring in the statistical analysis of probabilistic functions of Markov chains. The Annals of Mathematical Statistics, 41(1):164- 171.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "An inequality and associated maximization technique in statistical estimation of probabilistic functions of a Markov process",
"authors": [
{
"first": "L",
"middle": [
"E"
],
"last": "Baum",
"suffix": ""
}
],
"year": 1972,
"venue": "Inequalities",
"volume": "3",
"issue": "",
"pages": "1--8",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "L. E. Baum. 1972. An inequality and associated maxi- mization technique in statistical estimation of proba- bilistic functions of a Markov process. Inequalities, 3:1-8.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "A gentle tutorial of the EM algorithm and its application to parameter estimation for Gaussian mixture and hidden Markov models",
"authors": [
{
"first": "J",
"middle": [],
"last": "Bilmes",
"suffix": ""
}
],
"year": 1997,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "J. Bilmes. 1997. A gentle tutorial of the EM algorithm and its application to parameter estimation for Gaus- sian mixture and hidden Markov models. Technical Report TR-97-021, ICSI.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Coupled hidden Markov models for complex action recognition",
"authors": [
{
"first": "M",
"middle": [],
"last": "Brand",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Oliver",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Pentland",
"suffix": ""
}
],
"year": 1997,
"venue": "Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition",
"volume": "",
"issue": "",
"pages": "994--999",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "M. Brand, N. Oliver, and A. Pentland. 1997. Coupled hidden Markov models for complex action recog- nition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 994-999.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Learning to sportscast: A test of grounded language acquisition",
"authors": [
{
"first": "D",
"middle": [
"L"
],
"last": "Chen",
"suffix": ""
},
{
"first": "R",
"middle": [
"J"
],
"last": "Mooney",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of the 25th International Conference on Machine Learning",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "D. L. Chen and R. J. Mooney. 2008. Learning to sportscast: A test of grounded language acquisition. In Proceedings of the 25th International Conference on Machine Learning.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "On nonlinear fractional programming",
"authors": [
{
"first": "W",
"middle": [],
"last": "Dinkelbach",
"suffix": ""
}
],
"year": 1967,
"venue": "Management Science",
"volume": "13",
"issue": "7",
"pages": "492--498",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "W. Dinkelbach. 1967. On nonlinear fractional pro- gramming. Management Science, 13(7):492-498.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Learning to talk about events from narrated video in a construction grammar framework",
"authors": [
{
"first": "P",
"middle": [
"F"
],
"last": "Dominey",
"suffix": ""
},
{
"first": "J.-D",
"middle": [],
"last": "Boucher",
"suffix": ""
}
],
"year": 2005,
"venue": "Artificial Intelligence",
"volume": "167",
"issue": "12",
"pages": "31--61",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "P. F. Dominey and J.-D. Boucher. 2005. Learning to talk about events from narrated video in a construc- tion grammar framework. Artificial Intelligence, 167(12):31-61.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Object detection with discriminatively trained part-based models",
"authors": [
{
"first": "P",
"middle": [
"F"
],
"last": "Felzenszwalb",
"suffix": ""
},
{
"first": "R",
"middle": [
"B"
],
"last": "Girshick",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Mcallester",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Ramanan",
"suffix": ""
}
],
"year": 2010,
"venue": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"volume": "32",
"issue": "9",
"pages": "1627--1645",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "P. F. Felzenszwalb, R. B. Girshick, D. McAllester, and D. Ramanan. 2010a. Object detection with discrim- inatively trained part-based models. IEEE Transac- tions on Pattern Analysis and Machine Intelligence, 32(9):1627-1645.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Cascade object detection with deformable part models",
"authors": [
{
"first": "P",
"middle": [
"F"
],
"last": "Felzenszwalb",
"suffix": ""
},
{
"first": "R",
"middle": [
"B"
],
"last": "Girshick",
"suffix": ""
},
{
"first": "D",
"middle": [
"A"
],
"last": "Mcallester",
"suffix": ""
}
],
"year": 2010,
"venue": "Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition",
"volume": "",
"issue": "",
"pages": "2241--2248",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "P. F. Felzenszwalb, R. B. Girshick, and D. A. McAllester. 2010b. Cascade object detection with deformable part models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recog- nition, pages 2241-2248.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Baby talk: Understanding and generating simple image descriptions",
"authors": [
{
"first": "G",
"middle": [],
"last": "Kulkarni",
"suffix": ""
},
{
"first": "V",
"middle": [],
"last": "Premraj",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Dhar",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Choi",
"suffix": ""
},
{
"first": "A",
"middle": [
"C"
],
"last": "Berg",
"suffix": ""
},
{
"first": "T",
"middle": [
"L"
],
"last": "Berg",
"suffix": ""
}
],
"year": 2011,
"venue": "Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition",
"volume": "",
"issue": "",
"pages": "1601--1608",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "G. Kulkarni, V. Premraj, S. Dhar, S. Li, Y. Choi, A. C. Berg, and T. L. Berg. 2011. Baby talk: Understand- ing and generating simple image descriptions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1601-1608.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Collective generation of natural image descriptions",
"authors": [
{
"first": "P",
"middle": [],
"last": "Kuznetsova",
"suffix": ""
},
{
"first": "V",
"middle": [],
"last": "Ordonez",
"suffix": ""
},
{
"first": "A",
"middle": [
"C"
],
"last": "Berg",
"suffix": ""
},
{
"first": "T",
"middle": [
"L"
],
"last": "Berg",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Choi",
"suffix": ""
}
],
"year": 2012,
"venue": "Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics: Long Papers",
"volume": "1",
"issue": "",
"pages": "359--368",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "P. Kuznetsova, V. Ordonez, A. C. Berg, T. L. Berg, and Y. Choi. 2012. Collective generation of natural im- age descriptions. In Proceedings of the 50th Annual Meeting of the Association for Computational Lin- guistics: Long Papers -Volume 1, pages 359-368.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "A probabilistic model of syntactic and semantic acquisition from child-directed utterances and their meanings",
"authors": [
{
"first": "T",
"middle": [],
"last": "Kwiatkowski",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Goldwater",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Zettlemoyer",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Steedman",
"suffix": ""
}
],
"year": 2012,
"venue": "Proceedings of the 13th Conference of the European Chapter of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "234--244",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "T. Kwiatkowski, S. Goldwater, L. Zettlemoyer, and M. Steedman. 2012. A probabilistic model of syn- tactic and semantic acquisition from child-directed utterances and their meanings. In Proceedings of the 13th Conference of the European Chapter of the As- sociation for Computational Linguistics, pages 234- 244.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Im2text: Describing images using 1 million captioned photographs",
"authors": [
{
"first": "V",
"middle": [],
"last": "Ordonez",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Kulkarni",
"suffix": ""
},
{
"first": "T",
"middle": [
"L"
],
"last": "Berg",
"suffix": ""
}
],
"year": 2011,
"venue": "Proceedings of Neural Information Processing Systems",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "V. Ordonez, G. Kulkarni, and T. L. Berg. 2011. Im2text: Describing images using 1 million cap- tioned photographs. In Proceedings of Neural In- formation Processing Systems.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Learning visually-grounded words and syntax for a scene description task",
"authors": [
{
"first": "D",
"middle": [],
"last": "Roy",
"suffix": ""
}
],
"year": 2002,
"venue": "Computer Speech and Language",
"volume": "16",
"issue": "",
"pages": "353--385",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "D. Roy. 2002. Learning visually-grounded words and syntax for a scene description task. Computer Speech and Language, 16:353-385.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Action bank: A high-level representation of activity in video",
"authors": [
{
"first": "S",
"middle": [],
"last": "Sadanand",
"suffix": ""
},
{
"first": "J",
"middle": [
"J"
],
"last": "Corso",
"suffix": ""
}
],
"year": 2012,
"venue": "Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition",
"volume": "",
"issue": "",
"pages": "1234--1241",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "S. Sadanand and J. J. Corso. 2012. Action bank: A high-level representation of activity in video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1234-1241.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "A maximumlikelihood approach to visual event classification",
"authors": [
{
"first": "J",
"middle": [
"M"
],
"last": "Siskind",
"suffix": ""
},
{
"first": "Q",
"middle": [],
"last": "Morris",
"suffix": ""
}
],
"year": 1996,
"venue": "Proceedings of the Fourth European Conference on Computer Vision",
"volume": "",
"issue": "",
"pages": "347--360",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "J. M. Siskind and Q. Morris. 1996. A maximum- likelihood approach to visual event classification. In Proceedings of the Fourth European Conference on Computer Vision, pages 347-360.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "A computational study of crosssituational techniques for learning word-to-meaning mappings",
"authors": [
{
"first": "J",
"middle": [
"M"
],
"last": "Siskind",
"suffix": ""
}
],
"year": 1996,
"venue": "Cognition",
"volume": "61",
"issue": "",
"pages": "39--91",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "J. M. Siskind. 1996. A computational study of cross- situational techniques for learning word-to-meaning mappings. Cognition, 61:39-91.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Realtime American Sign Language recognition using desk and wearable computer based video",
"authors": [
{
"first": "T",
"middle": [],
"last": "Starner",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Weaver",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Pentland",
"suffix": ""
}
],
"year": 1998,
"venue": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"volume": "20",
"issue": "12",
"pages": "1371--1375",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "T. Starner, J. Weaver, and A. Pentland. 1998. Real- time American Sign Language recognition using desk and wearable computer based video. IEEE Transactions on Pattern Analysis and Machine In- telligence, 20(12):1371-1375.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Error bounds for convolutional codes and an asymtotically optimum decoding algorithm",
"authors": [
{
"first": "A",
"middle": [
"J"
],
"last": "Viterbi",
"suffix": ""
}
],
"year": 1967,
"venue": "IEEE Transactions on Information Theory",
"volume": "13",
"issue": "",
"pages": "260--267",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "A. J. Viterbi. 1967. Error bounds for convolutional codes and an asymtotically optimum decoding algo- rithm. IEEE Transactions on Information Theory, 13:260-267.",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "Convolutional codes and their performance in communication systems",
"authors": [
{
"first": "A",
"middle": [],
"last": "Viterbi",
"suffix": ""
}
],
"year": 1971,
"venue": "IEEE Transactions on Communication Technology",
"volume": "19",
"issue": "5",
"pages": "751--772",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "A. Viterbi. 1971. Convolutional codes and their per- formance in communication systems. IEEE Trans- actions on Communication Technology, 19(5):751- 772.",
"links": null
},
"BIBREF23": {
"ref_id": "b23",
"title": "Recognizing human action in time-sequential images using hidden Markov model",
"authors": [
{
"first": "J",
"middle": [],
"last": "Yamoto",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Ohya",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Ishii",
"suffix": ""
}
],
"year": 1992,
"venue": "Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition",
"volume": "",
"issue": "",
"pages": "379--385",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "J. Yamoto, J. Ohya, and K. Ishii. 1992. Recogniz- ing human action in time-sequential images using hidden Markov model. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recog- nition, pages 379-385.",
"links": null
},
"BIBREF24": {
"ref_id": "b24",
"title": "I2T: Image parsing to text description. Proceedings of the IEEE",
"authors": [
{
"first": "B",
"middle": [
"Z"
],
"last": "Yao",
"suffix": ""
},
{
"first": "X",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Lin",
"suffix": ""
},
{
"first": "M",
"middle": [
"W"
],
"last": "Lee",
"suffix": ""
},
{
"first": "S.-C",
"middle": [],
"last": "Zhu",
"suffix": ""
}
],
"year": 2010,
"venue": "",
"volume": "98",
"issue": "",
"pages": "1485--1508",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "B. Z. Yao, X. Yang, L. Lin, M. W. Lee, and S.-C. Zhu. 2010. I2T: Image parsing to text description. Pro- ceedings of the IEEE, 98(8):1485-1508, August.",
"links": null
},
"BIBREF25": {
"ref_id": "b25",
"title": "On the integration of grounding language and learning objects",
"authors": [
{
"first": "C",
"middle": [],
"last": "Yu",
"suffix": ""
},
{
"first": "D",
"middle": [
"H"
],
"last": "Ballard",
"suffix": ""
}
],
"year": 2004,
"venue": "Proceedings of the 19th National Conference on Artifical intelligence",
"volume": "",
"issue": "",
"pages": "488--493",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "C. Yu and D. H. Ballard. 2004. On the integration of grounding language and learning objects. In Pro- ceedings of the 19th National Conference on Artifi- cal intelligence, pages 488-493.",
"links": null
},
"BIBREF26": {
"ref_id": "b26",
"title": "A new formulation of coupled hidden Markov models",
"authors": [
{
"first": "S",
"middle": [],
"last": "Zhong",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Ghosh",
"suffix": ""
}
],
"year": 2001,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "S. Zhong and J. Ghosh. 2001. A new formulation of coupled hidden Markov models. Technical report, Department of Electrical and Computer Engineer- ing, The University of Texas at Austin.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"text": "the left of the backpack carried the trash-can towards the chair.",
"num": null,
"type_str": "figure",
"uris": null
},
"FIGREF2": {
"text": "ROC curves of trained models and handwritten models.",
"num": null,
"type_str": "figure",
"uris": null
},
"TABREF1": {
"num": null,
"type_str": "table",
"content": "<table><tr><td>: Arguments and model configurations for</td></tr><tr><td>different parts of speech c. VEL stands for veloc-</td></tr><tr><td>ity, MAG for magnitude, ORIENT for orientation,</td></tr><tr><td>and DIST for distance.</td></tr></table>",
"text": "",
"html": null
},
"TABREF3": {
"num": null,
"type_str": "table",
"content": "<table/>",
"text": "F1 scores of different methods.",
"html": null
}
}
}
} |