Share
Embed
-SSII技術マップを通して見る過去・現在，そして未来-
SSII30周年記念パネルディスカッション
2024年6月12日（水）
https://confit.atlas.jp/guide/event/ssii2024/static/specialsession
Hironobu Fujiyoshi
June 12, 2024
Transcript
16 slides
-
ࣗݾհɿ౻٢߂ 2 த෦େֶϩΰ த෦େֶϩΰ ֶྺɿ த෦େֶେֶӃम࢜՝ఔमྃ த෦େֶେֶӃത࢜ޙظ՝ఔຬظୀֶʢത࢜ʣ ݚڀ׆ಈɿ ถΧʔωΪʔϝϩϯେֶϩϘοτֶݚڀॴϙευΫݚڀһʢʣ
த෦େֶֶ෦ߨࢣ த෦େֶֶ෦।ڭत ถΧʔωΪʔϝϩϯେֶϩϘοτֶݚڀॴ٬һݚڀһʢʣ த෦େֶֶ෦ڭत ～ 44**ϓϩάϥϜҕһ 44**࣮ߦҕһ 44**࣮ߦҕһɼػց֮ϩϘςΟΫεݚڀάϧʔϓ ݱࡏʹࢸΔ ֶ֎׆ಈɿ ຊσΟʔϓϥʔχϯάڠձཧࣄ ΫϩεΞϙΠϯτϝϯτʢσϯιʔʣ vol.162 ౻٢߂ʢத෦େֶʣʮ“+AI”ͰมΘΔະདྷʯ IUUQTXXXZPVUVCFDPNXBUDI W [K&0J7)6 ౻٢߂ ాӯࣿ͞Μ (೫ࡔ46) ాेເ͞Μ -
w 44**ͷपاը ਐలͷ֮·͍͠ը૾ηϯγϯ グ ٕज़Λཧ͠ɼ͔ͦ͜Β֎ૠͯ͠ະདྷΛ プ Ϩ ビ ϡʔ
44**ٕज़Ϛοϓ 2000 2005 2010 2015 幾 何 情 報 取 得 光 情 報 取 得 基 本 機 能 向 上 ( ) (PS ) / 光 取得 能( ) / (( ) 機 Pulse count Imspector(Specim, 99) / Entertainment Vision Sensor(SONY, 01/02) Safety Laser Scanner (SICK, 02) 光 TOF) Photon Ics (PMDTech, 03) 光 TOF)向 2 03 Equiox (Canesta, 04) 光 TOF Cartesia(Space Vision, 04) PMD(98) Pulse count (01) (CMOS) (04) LED LIDAR(05) TOF(12) 能up MF-PD TOF(2014) / 向 (03) (05) Catadioptic Imaging(06 (09 (10 光 光 (91 (95) RF(94) (01 光 光 (03 LSI VGA 光 (04) 光 (04 (06 Temporal Dithering(08 向 Binary Dithering(13 2 (13) FB PS(03) PS(09) + (99) DFD Primal-Dual Coding(12) 1 光 LT情報取得 (12) 3 6 HDTV(04) 2 (08) Coded Aperture + (10 Non-scanning (04 向 Light Stage2(02) 楕円面ミラー＋プロジェクタ方式(07) 入射・出射方向のサンプリング簡略化 SVBRDF (13) BSSRDF (13) (05 (07 光 情報 1 取得 光 (01) / DR (05) 光 Panchromatic pixels(12) Bayar S/N向上 Jitter Camera(04) (11) Multi Camara Array(04 Pleoptic Camera(05 MLA (11) 光 AWARE Gigapixel Imaging(13) 光 光 (06 3500fps 512x512 Femto Photography(11 能 能 Coded Aperture(07/09) 光 Focal Sweep(08) 能( ) (12) ( ) HDL LIDAR(Velodyne, 07) 光 TOF Grambel Precision, 08) ProFusion(PointGray, 08 Kinect (Prime Sense, 09) CAMERACHIP(Omnivision, 08) Wavefront Coding Gelsight Benchtop Scanner(Gelsight, 10) Photometric Stereo D-Imager Panasonic, 10) 光 TOF Raytrix R11(Raytrix, 10) Lytro Light Field Camera(Lytro, 12) Kinect2 (Microsoft, 13) 光 TOF TVS 3.0(3 , 14) 6500 12100 1000 9800 (00) 光 (00) 16 Spatially Varying Pixel Exposures(00) 光 Wavefront Coding(95/03/08) 2014/05/06版 PGP (95) TOMBO(01 ಛ ந ग़ ύ λ ồ ϯ Ϛ ỽ ν ϯ ά ಛ ݕ ग़ ɾ ه ड़ ౷ ܭ త ֶ श ๏ ࠷ ۙ ୳ ࡧ SVM(95) Ϛʔδϯ࠷େԽ SIFT(99) εέʔϧෆม ಛݕग़ɾهड़ SURF(06) ੵը૾ ΞϧΰϦζϜʹΑΔߴԽ GPU SIFT(06) ϋʔυΣΞʹΑΔߴԽ FAST(06) ػցֶश ίʔφʔݕग़ BRIEF(10) ֶशແ͠ ϥϯμϜαϯϓϦϯά ORB(11) ڭࢣແֶ͠श D-BRIEF(12) ڭࢣ͋Γֶश Bin-Boost(13) ڭࢣ͋Γֶश Random Forests(01) ΞϯαϯϒϧֶशʴϥϯμϜֶश DPM(08) Ϟσϧͷׂ latent SVM ʹΑΔࣝผ Exemplar SVM(11) ࣄྫϕʔεͷ SVM WTA Hashing(11) ߴԽ HOG(05) ޯใ HLAC(88) ߴ࣍ͷࣗݾ૬ؔ CHLAC(04) HLAC ʹ࣌ؒ࣠ͷՃ Haar-like(01) box ϑΟϧλ DOT(10) ޯใͷςϯϓϨʔτϚονϯά VLAD(10) ؔ࿈͢Δ VW ͷಛྔΛ༻ Fisher Vector(07) ֬ີؔʹΑΔಛྔͷදݱ Deep Learning(08) ଟχϡʔϥϧωοτϫʔΫ දݱֶश ଟΫϥεࣝผ ສΧςΰϦ Crowdsourcing(13) ਓͷݟͷಋೖ ৄࡉը૾ࣝผ ॲ ཧ Ϩ ϕ ϧ ERT(06) RF ͷϥϯμϜੑΛ࠷େԽ Fern(06) RF ͷذ݅Λ֊Ͱ౷Ұ 5 ޙ texton(01) ϑΟϧλͷόϯΫ ਓମύʔπࣝผ ߴԽ ߴਫ਼Խ ΞϓϦέʔγϣϯ 2000 2005 2010 ମݕग़ ଟΫϥε ࣗݾҐஔਪఆ ϚγϯϏδϣϯ ը૾ݕࡧ إݕग़ ਓݕग़ ಛఆମೝࣝ ը૾ྨ ಛྔͷࣗಈੜ ਓ͕ͨ͠Ґஔ͔Βಛநग़ ϚʔΧೝࣝ ೋಛ ਓͱͷϋΠϒϦουʹΑΔ ೳݕࠪɺܽؕݕग़ ૿ූ߸૬ؔ (00) ًͷ૿ݮΛೋͰը૾Խ RRF(03) ํͷೱมԽ ૄςϯϓϨʔτϚονϯά (05) छྨͷϞσϧͷ͍͚ ݻ༗ղςϯϓϨʔτϚονϯά (11) ճసมԽʹؤ݈ͳใΛར༻ Co-Occurrence Template Matching(10) ݦஶੑͷߴ͍ըૉͰর߹ ߴଟΫϥεࣝผ Harris-Affine(02) ΞϑΟϯෆมಛݕग़ MSER(02) ߴͳΞϑΟϯෆมಛ Object Bank(10) ଟΫϥεͷཁૉΛಛྔԽ Relative attribute(11) ࣮ʹΑΔؔ࿈ཁૉͷදݱ zero-shot transfer(09) ؔ࿈ཁૉ͔ΒඇֶशΫϥεͷݕग़ zero-shot learning ʹΑΔ ֶश֎αϯϓϧͷదԠ సҠֶश ܭྔֶश ੜଶௐࠪ ηϚϯςΟοΫηάϝϯςʔγϣϯ CoHOG(09) )0( ͷڞىදݱ MLP(86) ଟύʔηϒτϩϯ Online PA(06) ೖྗαϯϓϧʹԠͯ͡ॏΈϕΫτϧߋ৽ εύʔεಛྔ (06) Haar-like ϐΫηϧࠩ ৄࡉը૾هड़ άϥεϚϯଟ༷ମ (08) ઢܗ෦ۭؒͷू߹ମ product quantization(11) αϒϕΫτϧʹΑΔྔࢠԽ εϖΫτϧཧʹΑΔεέʔϧ୳ࡧ (12) ಛۭؒͷࣹӨ ෆมੑͷ֫ಘ DAISY(08) هड़ۭؒͷվྑ BOF(04) ಛྔͷࣙॻԽ େنإೝࣝ Deep Neural Network ͷ ߴԽ ަ੍૬ޓۭؒ๏ (06) ަߦྻʹΑΔۭؒͷؔΛߦԽ ΧʔωϧτϦοΫ (00) ಛۭؒͷࣹӨ ݻ༗ۭؒ๏ (96) ࣍ݩը૾ʹΑΔ ࣍ݩମೝࣝ ૬ޓ෦ۭؒ (85) ෦ۭؒಉ࢜ͷਖ਼४֯ LBP(94) ہॴྖҬͷೋԽ CARD(11) ಛྔΛ Խ Decision Jungles(13) ύεڞ༗ʹΑΔলϝϞϦͳܾఆ ੍૬ޓ෦ۭؒ๏ (99) ࣝผʹ༗ޮͳۭؒͷࣹӨ Pentium3(99) Pentium4(00) Xeon5100 Intel Core2 Core i 7(11) CPU (06) Core i 5(09) Core i 3(10) SSE SSE(99) SSE2(00) SSE3(04) SSE4.1(08) SSE4.2(11) GPU GeForce2(00) GeForce3(01) GeForce4,FX(02) GeForce6(04) GeForce7(05) GeForce8(06) GeForce9(08) GeForce200(08) GeForce400(10) GeForce500(11) GeForce600(12) GeForce700(13) CUDA Tesla(07) R100(00) R200,300(02) R400(04) Northern Islands(11) R500(05) R600(07) R700(08) Evergreen(10) Southern Islands(12) Volcanic Islands(13) 2014/06/09 ൛ CNN(89) ϓʔϦϯάͱΈࠐΈ ʹΑΔಛநग़ AdaBoost(95) Ξϯαϯϒϧֶश αϯϓϧॏΈͷஞ࣍ߋ৽ ηϚϯςΟοΫө૾ѹॖ KE` ÏßÙ°º« ¬¬ ćĔêĔüčùëĒíĕ êĉčlC+ 1 ¥À¹Ë½¶À¹ ¢µÀ½¶ÅµÇ½Ã ¥µ·ÇÃŽ͵ǽà®® ®Å½ÃŽǽ͹¸££ ÁµÇ·¼½Â» ¡È¸À¹ ¸¾ÈÆÇÁ¹ÂÇ «ÈÀǽÄÀ¹¶µÆ¹À½Â¹ ¯¹»ÈÀµÅ½ÍµÇ½Ã ®ÀµÂ¹Æʹ¹Ä ¸µÄǽɹ ʽ¸ÃÊ ±¹Çŵ¼¹¸Åµ·µÅɽ» £Ã¦ē°¨¥± §µÅÅ½Æ ¥ °± £ª± ®Åþ¹·Ç½É¹¥ £ ¨°´ ¢ÅÃÆƶµÆ¹¸À÷µÀ ÆÈÄÄÃÅÇ ¦ÀöµÀÁµÇ·¼½Â» íčĀêùü ¡Å½¹º®ÅÃĵ»µÇ½Ã ¬Ã®µÅµÁ¹ÇŽ· ¢µÀ½¶ÅµÇ½Ã °Äµ·¹·µÅɽ» ª© ®± « ¯ÇÃÃÀ¿½Ç Ė1ėØ~ÜØ7Î äďïĎôĈÐ h2µÂ¸ÃÅđþóüµÂ¸ÃÅ| ¡ ½ÂÀµÅ»¹ ĀèüĉüĎÖÙo ruOÙñĔĒ% N w 28 ğ "\uOÙ°º« đĔĎĒíñċùöĔ ÙwR ¡È½À¸½Â»¯ÃÁµ aÙčĒýćĔì ÏßÙlC+ *mûĔöÄŽÃÅÙ _Ėė 6ğ ĞÄÇ¥«µÇÅ½Ë ĜÄǤ«µÇÅ½Ë ÄÇ¥«µÇÅ½Ë ęɽ¹Ê°º« êĉč l + ¬É½¹Ê°º« Y9Nē y{ē ćùøĒí ðĔĀçó ^= êĉčMQ ąìõďeàÑ ěÕes[ĕI &vU;$Ù_ 0óëċĒØÞágTĕ| đþóü×y{(â_ îóüĖÛėâ:bØxf ®«³° QÙ_ ªÃ·µÀÆÈÄÄÃÅÇÙ_ îóüÙI °¹¹¸¦ÅÃÊăĔóÙ đþóüćùøĒí đĔêďÿùø °°£ XÚÙA3 ®Å¹¹ÁÄǽɹ ¯¤¢¬ ° ¢Ùj W ÙbR+ ĊûďĀåùúæĒíĕ&DR ,H]I ,HêĉčüčùëĒí ĕeU8 ` NiÖÙo ĕ xfÚÙI ÀµÅ»¹Æ·µÀ¹Õٮ® ¢¦ÄŹ·Ã¸½Ç½ÃµÅØÞá 'uO| êĉčpf êĉčlf êĉčlēeUI 0MQÿöĔĒâ_ÎÒ gGëċĎāĐĔñČĒ êĉčÿčĉĔöÙA3 £½Å¹·ÇµÀ½»ÂÁ¹ÂÇ £± « ³ÃÀÈÁ¹ÇŽ·»ÅµÄ¼·ÈÇÆ íčĀêùüÕI×ðĔĀçó@ ,H«³°ûĂóćùĂâ _ÑÒêĉčüčùëĒí ¨¢¨ ØÞá|ĕ` Õ ĐòóüĐĔñČĒ £¹Âƹ óúĐé ¿¸JăĔóÙ |Iz Bi Y9Ù? ©ª± ûĀãìüóöĒ÷Ĕý Ø×ÓÔÎáÝÙ ĐĒ÷ĎĒí ¢Ã¼¹Å¹ÂǽÁµ»¹Æ §Ì¶Å½¸Áø¹À½Â» !(wØÞá8 ĖäĀåĒêĉčĊûďė .5êĉčĊûďÚÙA3 vU )úìóøċćùĂ ĊûďăĔóýóúĐé ¨ÀÀÈÁ½ÂµÇ½Ã·Ã¹ čĒþĔüĕ<US VF` Ùk4 = &vU` «³°ĘUĄä Ïß.Zâxf ®¼ÃÇÃÇÃÈŽÆÁ ĚP 0zĊûďÞá 'ûĔöÙĐĒ÷ĎĒí « ³² ³ÚÙ:_ ½Á»ÄÆ Ė` ėNiÕ n/lC+ ¡È¸À¹Å «ÃÂ𪠫 cZbĚ£ćùĂs[ ĖêĉčĆĔôĕY9ė ¬ÃÂŽ»½¸°º« ¯¯° ¢ ¯ ¬° ¢ ¯ÃÀÀ½Â»°¼ÈÇǹŠ¢µÁ¹Åµ®Åþ¹·Ç½Ã #dpf ĝUÕ#dpf #Lpf ¢«°êĉč-: Tuesday, May 20, 14 Πϝʔ ジ ϯ グ ɹೝࣝɹ ࡾ࣍ݩγʔϯͷ෮ݩ ਡ๚ਖ਼थʢΦϜϩϯ 049ʣ ౻٢߂ʢத෦େֶʣ ௗډळʢ౦ژۀେֶʣ ը૾ηϯγϯάͱͯ͠ͷೖྗܥΛޫͷ࣋ͭ ༷ʑͳಛੑ͔Β၆ᛌ ಛɾهड़ɼಛྔɼࣝผثͷมભΛ·ͱΊɼ ͦͷಈΛհ δΦϝτϦɼϚονϯάɼαʔϑΣεੜͷมભΛ ·ͱΊɼͦͷಈΛհ Ӝ৻࡞ʢฌݿݝཱେʣ -
1 SSII2014「イメージング」 技術マップの振り返り １．イメージング技術分類の３本柱（幾何学情報取得/光情報取得/基本性能向上）について l ToF, LiDARや偏光カメラ等の認知・利⽤の広がり →ToF/LiDARは、完全に産業化・社会実装フェーズに移⾏（ADAS, ⾃動⾛⾏、スマホ搭載等が牽引） →とくにADAS向けLiDARにおいては光効率を担保しつつさらなるローコスト化を進めるにあたり、従来のメカニカルスキャン型に代わる
ソリッドステート型への市場期待は2014年以降依然として⾼いものの普及には⾄らず l 幾何学情報・光情報取得で活⽤される投光制御デバイス（プロジェクタ）の普及が急加速 →当時はDLP（DMD）⽅式の孤軍奮闘を想定していたが、LCOS（反射型液晶）⽅式の映画館やシアター採⽤が起爆剤となり、 ファクトリーオートメーション⽤途（検査装置や計測センサなど）への展開・普及が進む l 基本性能向上における「撮像素⼦の⾼解像度化にともなうリソース再配分」については、ライトフィールドカメラなど特化型製品としての 市場形成には⾄らず ２．新動向 l 超⾼分解能の時刻情報（SPAD, イベントカメラ）を距離計測以外に応⽤する研究の進展 →NLOS（Non-line-of-sight︓死⾓を観測する技術） →ToFヒストグラムやイベントフレームなどのフレーム時間内積分を⾏わない処理法の追求 l イメージング対象の拡⼤ →⾳場のイメージング（偏光カメラや⾼速カメラの活⽤）や触覚のイメージング →触覚のイメージングについては視触覚（Visuotactile）センサのロボットハンド制御での認知・利⽤が拡⼤ 44**ٕज़Ϛοϓ ʮΠϝʔδϯάʯͷৼΓฦΓ ਡ๚ਖ਼थʢΦϜϩϯ 049ʣ Ӝ৻࡞ʢฌݿݝཱେʣ CONFIDENTIAL EXTENDED ABSTRACT. DO NOT DISTRIBUTE ANYWHERE. ୈ 26 ճ ը૾ͷೝࣝɾཧղγϯϙδϜ NLOS-NeuS: ඇࢹઢํࡱӨʹ͓͚Δ χϡʔϥϧӄؔද໘ ౻ଜ ༑و1,a) ᷱా و߂1,b) ᢠ ࠸1,c) ߁ത1,d) ֓ཁ ඇࢹઢํࡱӨʢNon-line-of-sight (NLOS) imagingʣͱ ɼΧϝϥ͔Βݟ͑ͳ͍γʔϯΛ֮͢Δٕज़Ͱ͋Δɽຊ ݚڀͰߴ࣌ؒղσʔλΛೖྗͱ͠ɼ χϡʔϥϧӄؔද ໘ʹΑΔ NLOS γʔϯͷ 3 ࣍ݩܗঢ়෮ݩख๏ΛఏҊ͢Δɽ χϡʔϥϧӄؔද໘ͱͯ͠ූ߸͖ڑؔʢSigned distance function (SDF)ʣΛଟύʔηϓτϩϯͰදݱ͠ɼ NLOS γʔϯʹ͓͍ͯ SDF Λਖ਼ֶ͘͠श͢ΔͨΊͷ੍ Λಋೖ͢ΔɽϘΫηϧදݱΛ༻͍ͨैདྷख๏ʹൺͯɼৄ ࡉ͔ͭΒ͔ͳܗঢ়෮ݩ͕ՄೳͰ͋Δ͜ͱΛࣔ͢ɽ 1. ͡Ίʹ ۙίϯϐϡʔλϏδϣϯͷͰɼ ଟ͘ͷΞϓϦέʔ Relay wall NLOS scene Collocated light source and detector (a) NLOS setup f photons /POMJOFPGTJHIU /-04 JNBHJOH -
ը૾ೝࣝͷظͱՄೳੑɹத෦େֶֶ෦ใֶՊɹ౻٢߂ 2 ʮೝࣝʯͰऔΓѻ͏ٕज़ྖҬ ಛ ந ग़ ύ λ ồ
ϯ Ϛ ỽ ν ϯ ά ಛ ݕ ग़ ɾ ه ड़ ౷ ܭ త ֶ श ๏ ࠷ ۙ ୳ ࡧ SVM(95) Ϛʔδϯ࠷େԽ SIFT(99) εέʔϧෆม ಛݕग़ɾهड़ SURF(06) ੵը૾ ΞϧΰϦζϜʹΑΔߴԽ GPU SIFT(06) ϋʔυΣΞʹΑΔߴԽ FAST(06) ػցֶश ίʔφʔݕग़ BRIEF(10) ֶशແ͠ ϥϯμϜαϯϓϦϯά ORB(11) ڭࢣແֶ͠श D-BRIEF(12) ڭࢣ͋Γֶश Bin-Boost(13) ڭࢣ͋Γֶश Random Forests(01) ΞϯαϯϒϧֶशʴϥϯμϜֶश DPM(08) Ϟσϧͷׂ latent SVM ʹΑΔࣝผ Exemplar SVM(11) ࣄྫϕʔεͷ SVM WTA Hashing(11) ߴԽ HOG(05) ޯใ HLAC(88) ߴ࣍ͷࣗݾ૬ؔ CHLAC(04) HLAC ʹ࣌ؒ࣠ͷՃ Haar-like(01) box ϑΟϧλ DOT(10) ޯใͷςϯϓϨʔτϚονϯά VLAD(10) ؔ࿈͢Δ VW ͷಛྔΛ༻ Fisher Vector(07) ֬ີؔʹΑΔಛྔͷදݱ Deep Learning(08) ଟχϡʔϥϧωοτϫʔΫ දݱֶश ଟΫϥεࣝผ ສΧςΰϦ Crowdsourcing(13) ਓͷݟͷಋೖ ৄࡉը૾ࣝผ ERT(06) RF ͷϥϯμϜੑΛ࠷େԽ Fern(06) RF ͷذ݅Λ֊Ͱ౷Ұ texton(01) ϑΟϧλͷόϯΫ ਓମύʔπࣝผ ߴԽ ߴਫ਼Խ ΞϓϦέʔγϣϯ 2000 2005 2010 ମݕग़ ଟΫϥε ࣗݾҐஔਪఆ ϚγϯϏδϣϯ ը૾ݕࡧ إݕग़ ਓݕग़ ಛఆମೝࣝ ը૾ྨ ಛྔͷࣗಈੜ ਓ͕ͨ͠Ґஔ͔Βಛநग़ ϚʔΧೝࣝ ೋಛ ૿ූ߸૬ؔ (00) ًͷ૿ݮΛೋͰը૾Խ RRF(03) ํͷೱมԽ ૄςϯϓϨʔτϚονϯά (05) छྨͷϞσϧͷ͍͚ ݻ༗ղςϯϓϨʔτϚονϯά (11) ճసมԽʹؤ݈ͳใΛར༻ Co-Occurrence Template Matching(10) ݦஶੑͷߴ͍ըૉͰর߹ Harris-Affine(02) ΞϑΟϯෆมಛݕग़ MSER(02) ߴͳΞϑΟϯෆมಛ Object Bank(10) ଟΫϥεͷཁૉΛಛྔԽ Relative attribute(11) ࣮ʹΑΔؔ࿈ཁૉͷදݱ zero-shot transfer(09) ؔ࿈ཁૉ͔ΒඇֶशΫϥεͷݕग़ ηϚϯςΟοΫηάϝϯςʔγϣϯ CoHOG(09) )0( ͷڞىදݱ MLP(86) ଟύʔηϒτϩϯ Online PA(06) ೖྗαϯϓϧʹԠͯ͡ॏΈϕΫτϧߋ৽ εύʔεಛྔ (06) Haar-like ϐΫηϧࠩ άϥεϚϯଟ༷ମ (08) ઢܗ෦ۭؒͷू߹ମ product quantization(11) αϒϕΫτϧʹΑΔྔࢠԽ εϖΫτϧཧʹΑΔεέʔϧ୳ࡧ (12) ಛۭؒͷࣹӨ DAISY(08) هड़ۭؒͷվྑ BOF(04) ಛྔͷࣙॻԽ ަ੍૬ޓۭؒ๏ (06) ަߦྻʹΑΔۭؒͷؔΛߦԽ ΧʔωϧτϦοΫ (00) ಛۭؒͷࣹӨ ݻ༗ۭؒ๏ (96) ࣍ݩը૾ʹΑΔ ࣍ݩମೝࣝ ૬ޓ෦ۭؒ (85) ෦ۭؒಉ࢜ͷਖ਼४֯ LBP(94) ہॴྖҬͷೋԽ CARD(11) ಛྔΛ Խ Decision Jungles(13) ύεڞ༗ʹΑΔলϝϞϦͳܾఆ ੍૬ޓ෦ۭؒ๏ (99) ࣝผʹ༗ޮͳۭؒͷࣹӨ CNN(89) ϓʔϦϯάͱΈࠐΈ ʹΑΔಛநग़ AdaBoost(95) Ξϯαϯϒϧֶश αϯϓϧॏΈͷஞ࣍ߋ৽ w ಛݕग़ɾهड़ͷಈ 44**ٕज़Ϛοϓ ʮೝࣝʯͷৼΓฦΓ ը૾ೝࣝͷظͱՄೳੑɹத෦େֶֶ෦ใֶՊɹ౻٢߂ 1 ʮೝࣝʯͰऔΓѻ͏ٕज़ྖҬ ಛݕग़ɾهड़ͷಈ ಛ ந ग़ ύ λ ồ ϯ Ϛ ỽ ν ϯ ά ಛ ݕ ग़ ɾ ه ड़ ౷ ܭ త ֶ श ๏ ࠷ ۙ ୳ ࡧ SIFT(99) εέʔϧෆม ಛݕग़ɾهड़ SURF(06) ੵը૾ ΞϧΰϦζϜʹΑΔߴԽ GPU SIFT(06) ϋʔυΣΞʹΑΔߴԽ FAST(06) ػցֶश ίʔφʔݕग़ BRIEF(10) ֶशແ͠ ϥϯμϜαϯϓϦϯά ORB(11) ڭࢣແֶ͠श D-BRIEF(12) ڭࢣ͋Γֶश Bin-Boost(13) ڭࢣ͋Γֶश 2000 2005 2010 ը૾ݕࡧ Harris-Affine(02) ΞϑΟϯෆมಛݕग़ MSER(02) ߴͳΞϑΟϯෆมಛ εϖΫτϧཧʹΑΔεέʔϧ୳ࡧ (12) ಛۭؒͷࣹӨ DAISY(08) هड़ۭؒͷվྑ CARD(11) ಛྔΛ Խ DOT(10) ޯใͷςϯϓϨʔτϚονϯά ೋಛ ૿ූ߸૬ؔ (00) ًͷ૿ݮΛೋͰը૾Խ RRF(03) ํͷೱมԽ ૄςϯϓϨʔτϚονϯά (05) छྨͷϞσϧͷ͍͚ HOG(05) ޯใ HLAC(88) ߴ࣍ͷࣗݾ૬ؔ CHLAC(04) HLAC ʹ࣌ؒ࣠ͷՃ Haar-like(01) box ϑΟϧλ VLAD(10) ؔ࿈͢Δ VW ͷಛྔΛ༻ Fisher Vector(07) ֬ີؔʹΑΔಛྔͷදݱ Crowdsourcing(13) ਓͷݟͷಋೖ ৄࡉը૾ࣝผ texton(01) ϑΟϧλͷόϯΫ ϚγϯϏδϣϯ إݕग़ ਓݕग़ ಛఆମೝࣝ ը૾ྨ ಛྔͷࣗಈੜ ਓ͕ͨ͠Ґஔ͔Βಛநग़ ϚʔΧೝࣝ ݻ༗ղςϯϓϨʔτϚονϯά (11) ճసมԽʹؤ݈ͳใΛར༻ Co-Occurrence Template Matching(10) ݦஶੑͷߴ͍ըૉͰর߹ Object Bank(10) ଟΫϥεͷཁૉΛಛྔԽ CoHOG(09) )0( ͷڞىදݱ εύʔεಛྔ (06) Haar-like ϐΫηϧࠩ BOF(04) ಛྔͷࣙॻԽ ަ੍૬ޓۭؒ๏ (06) ަߦྻʹΑΔۭؒͷؔΛߦԽ ΧʔωϧτϦοΫ (00) ಛۭؒͷࣹӨ ݻ༗ۭؒ๏ (96) ࣍ݩը૾ʹΑΔ ࣍ݩମೝࣝ ૬ޓ෦ۭؒ (85) ෦ۭؒಉ࢜ͷਖ਼४֯ LBP(94) ہॴྖҬͷೋԽ ੍૬ޓ෦ۭؒ๏ (99) ࣝผʹ༗ޮͳۭؒͷࣹӨ SVM(95) Ϛʔδϯ࠷େԽ Random Forests(01) ΞϯαϯϒϧֶशʴϥϯμϜֶश DPM(08) Ϟσϧͷׂ latent SVM ʹΑΔࣝผ Exemplar SVM(11) ࣄྫϕʔεͷ SVM WTA Hashing(11) ߴԽ Deep Learning(08) ଟχϡʔϥϧωοτϫʔΫ දݱֶश ଟΫϥεࣝผ ສΧςΰϦ ERT(06) RF ͷϥϯμϜੑΛ࠷େԽ Fern(06) RF ͷذ݅Λ֊Ͱ౷Ұ ਓମύʔπࣝผ ମݕग़ ଟΫϥε ࣗݾҐஔਪఆ Relative attribute(11) ࣮ʹΑΔؔ࿈ཁૉͷදݱ zero-shot transfer(09) ؔ࿈ཁૉ͔ΒඇֶशΫϥεͷݕग़ ηϚϯςΟοΫηάϝϯςʔγϣϯ MLP(86) ଟύʔηϒτϩϯ Online PA(06) ೖྗαϯϓϧʹԠͯ͡ॏΈϕΫτϧߋ৽ άϥεϚϯଟ༷ମ (08) ઢܗ෦ۭؒͷू߹ମ product quantization(11) αϒϕΫτϧʹΑΔྔࢠԽ Decision Jungles(13) ύεڞ༗ʹΑΔলϝϞϦͳܾఆ CNN(89) ϓʔϦϯάͱΈࠐΈ ʹΑΔಛநग़ AdaBoost(95) Ξϯαϯϒϧֶश αϯϓϧॏΈͷஞ࣍ߋ৽ 4*'5ͷߴԽ ը૾ೝࣝͷظͱՄೳੑɹத෦େֶֶ෦ใֶՊɹ౻٢߂ 7 ಛݕग़ͷߴԽ εέʔϧεϖʔεͷߴԽ ܾఆʹΑΔߴԽ ɾ463' ੵը૾Λ༻͍ͨۙࣅϔοηߦྻʹΑΔߴͳΩʔϙΠϯτݕग़ ɾ'"45 ػցֶशʢܾఆʣΛಋೖͯ͠ίʔφʔݕग़ΛߴԽ ɾεϖΫτϧཧ εϖΫτϧཧʹΑΔεέʔϧ୳ࡧͷߴԽͱߴਫ਼Խ ը૾ೝࣝͷظͱՄೳੑɹத෦େֶֶ෦ใֶՊɹ౻٢߂ 9 ಛهड़ͷߴԽ ̎ಛྔͷಋೖ ɾ#3*&' ɼ03# $"3% ڑܭࢉΛߟྀͨ͠2ʹΑΔಛهड़ ɾ%#3*&' #JO#PPTU ڭࢣ͋ΓֶशʹΑΔ࠷దͳ2ύλʔϯͷ֫ಘ ϙΠϯτ ɹಛྉΛʹ͢Δ͜ͱͰڑܭࢉʢϋϛϯάڑʣΛߴԽɺ44&ͷར༻ ɹলϝϞϦԽಉ࣌ʹ࣮ݱ ϙδςΟϒαϯϓϧ ωΨςΟϒαϯϓϧ 03#ͷࢀরϖΞ %#3*&'ʹ͓͚Δڭࢣ͋Γֶश OS2Ͱؔ࿈ߨԋ͋Γʂ ʢମೝࣝͷ࠷લઢʣ ը૾ೝࣝͷظͱՄೳੑɹத෦େֶֶ෦ใֶՊɹ౻٢߂ w %JTUJODUJWFJNBHFGFBUVSFTGSPNTDBMFJOWBSJBOULFZQPJOUT 4*'5 <-PXF> r εέʔϧɾճసʹෆมͳಛݕग़ɾهड़ 6 ಛݕग़ɾهड़ ϙΠϯτ ɹ%0( %JGGFSFODFPG(BVTTJBO ʹΑΔΩʔϙΠϯτݕग़ ɹޯํώετάϥϜʹΑΔಛهड़ %P(ը૾ ฏԽը૾ € σ0 € kσ0 € k2σ0 εέʔϧ € k3σ0 k⒢М⒨ - - - - ը૾ೝࣝͷظͱՄೳੑɹத෦େֶֶ෦ใֶՊɹ౻٢߂ w %JTUJODUJWFJNBHFGFBUVSFTGSPNTDBMFJOWBSJBOULFZQPJOUT 4*'5 <-PXF> r εέʔϧɾճసʹෆมͳಛݕग़ɾهड़ 8 ಛݕग़ɾهड़ ϙΠϯτ ɹ%0( %JGGFSFODFPG(BVTTJBO ʹΑΔΩʔϙΠϯτݕग़ ɹޯํώετάϥϜʹΑΔಛهड़ ׂ ׂ ํ Ψε૭ -
w ہॴಛྔɾ౷ܭతֶश๏ͷಈ ը૾ೝࣝͷظͱՄೳੑɹத෦େֶֶ෦ใֶՊɹ౻٢߂ 11 ಛநग़ͱ౷ܭతֶश๏ ہॴը૾ಛྔͱ౷ܭతֶश๏ ɾإݕग़ˠ)BBSMJLF εύʔεಛ
ʴ"EB#PPTU ɾาߦऀݕग़ˠ)0( 47. ϙΠϯτ ɹઃఆʹ߹Θͤͯಛྔ )BOEDSBGUFEGFBUVSF Λઃܭ ɹΫϥε͔ΒଟΫϥε ɾਓମύʔπࣝผˠ3BOEPN'PSFTU ɾը૾ྨˠ4*'5 #0' ʴ47. ը૾ೝࣝͷظͱՄೳੑɹத෦େֶֶ෦ใֶՊɹ౻٢߂ w "%JTDSJNJOBUJWFMZ5SBJOFE .VMUJTDBMF %FGPSNBCMF1BSU.PEFM<'FM[FOT[XBMC> r -BUFOU47.Λ༻͍ͨύʔπϕʔεͷମݕग़ 12 %1.ɿύʔπϕʔεͷମݕग़ ϙΠϯτ ɹମΛύʔπͷू߹ͱͯ͠දݱ %FGPSNBCMF1BSUT.PEFM ɹύʔπͷҐஔؔΛߟྀ͢Δ͜ͱͰ࢟มಈʹରԠ ϧʔτϑΟϧλ ύʔπϑΟϧλ ύʔπϑΟϧλͷ Ґஔؔ ը૾ೝࣝͷظͱՄೳੑɹத෦େֶֶ෦ใֶՊɹ౻٢߂ w 'BTU "DDVSBUF%FUFDUJPOPG 0CKFDU$MBTTFTPOB4JOHMF.BDIJOF<%FBO> r ສछྨͷମΛඵҎԼͰݕग़ 13 όΠφϦίʔυΛ༻͍ͨ)BTIʹΑΔສछྨͷମݕग़ Locality-sensitive Hashing with WTA WTA codeをP個に分割 P個のコードそれぞれの Hashテーブルを参照 クラス毎の スコアヒストグラムを作成 各クラスのフィルタ応 答マップを作る HOG特徴量 111101010011 WAT code ϙΠϯτ ɹଟΫϥε%1.ͷߴԽ ɹύʔπͷू߹ʹରͯ͠ɺ85")BTIΛར༻ͯ͠ଟΫϥεͷݕग़Λ࣮ݱ ը૾ೝࣝͷظͱՄೳੑɹத෦େֶֶ෦ใֶՊɹ౻٢߂ 14 όΠφϦίʔυΛ༻͍ͨ)BTIʹΑΔສछྨͷମݕग़ w 'BTU "DDVSBUF%FUFDUJPOPG 0CKFDU$MBTTFTPOB4JOHMF.BDIJOF<%FBO> r ສछྨͷମΛඵҎԼͰݕग़ ը૾ೝࣝͷظͱՄೳੑɹத෦େֶֶ෦ใֶՊɹ౻٢߂ 2 ʮೝࣝʯͰऔΓѻ͏ٕज़ྖҬ ηϚϯςΟοΫηάϝϯςʔγϣϯ RRF(03) ํͷೱมԽ ૄςϯϓϨʔτϚονϯά (05) छྨͷϞσϧͷ͍͚ HLAC(88) ߴ࣍ͷࣗݾ૬ؔ ૿ූ߸૬ؔ (00) ًͷ૿ݮΛೋͰը૾Խ ΧʔωϧτϦοΫ (00) ಛۭؒͷࣹӨ ݻ༗ۭؒ๏ (96) ࣍ݩը૾ʹΑΔ ࣍ݩମೝࣝ ૬ޓ෦ۭؒ (85) ෦ۭؒಉ࢜ͷਖ਼४֯ LBP(94) ہॴྖҬͷೋԽ SIFT(99) εέʔϧෆม ಛݕग़ɾهड़ SURF(06) ੵը૾ ΞϧΰϦζϜʹΑΔߴԽ GPU SIFT(06) ϋʔυΣΞʹΑΔߴԽ FAST(06) ػցֶश ίʔφʔݕग़ BRIEF(10) ֶशແ͠ ϥϯμϜαϯϓϦϯά ORB(11) ڭࢣແֶ͠श D-BRIEF(12) ڭࢣ͋Γֶश Bin-Boost(13) ڭࢣ͋Γֶश ը૾ݕࡧ Harris-Affine(02) ΞϑΟϯෆมಛݕग़ MSER(02) ߴͳΞϑΟϯෆมಛ εϖΫτϧཧʹΑΔεέʔϧ୳ࡧ (12) ಛۭؒͷࣹӨ DAISY(08) هड़ۭؒͷվྑ DOT(10) ޯใͷςϯϓϨʔτϚονϯά VLAD(10) ؔ࿈͢Δ VW ͷಛྔΛ༻ Fisher Vector(07) ֬ີؔʹΑΔಛྔͷදݱ Crowdsourcing(13) ਓͷݟͷಋೖ ৄࡉը૾ࣝผ ϚγϯϏδϣϯ ಛఆମೝࣝ ಛྔͷࣗಈੜ ਓ͕ͨ͠Ґஔ͔Βಛநग़ ϚʔΧೝࣝ ೋಛ ݻ༗ղςϯϓϨʔτϚονϯά (11) ճసมԽʹؤ݈ͳใΛར༻ Co-Occurrence Template Matching(10) ݦஶੑͷߴ͍ըૉͰর߹ Object Bank(10) ଟΫϥεͷཁૉΛಛྔԽ CoHOG(09) )0( ͷڞىදݱ Online PA(06) ೖྗαϯϓϧʹԠͯ͡ॏΈϕΫτϧߋ৽ Exemplar SVM(11) ࣄྫϕʔεͷ SVM Deep Learning(08) ଟχϡʔϥϧωοτϫʔΫ දݱֶश ମݕग़ ଟΫϥε ࣗݾҐஔਪఆ Relative attribute(11) ࣮ʹΑΔؔ࿈ཁૉͷදݱ zero-shot transfer(09) ؔ࿈ཁૉ͔ΒඇֶशΫϥεͷݕग़ άϥεϚϯଟ༷ମ (08) ઢܗ෦ۭؒͷू߹ମ product quantization(11) αϒϕΫτϧʹΑΔྔࢠԽ WTA Hashing(11) ߴԽ ଟΫϥεࣝผ ສΧςΰϦ ERT(06) RF ͷϥϯμϜੑΛ࠷େԽ Fern(06) RF ͷذ݅Λ֊Ͱ౷Ұ Decision Jungles(13) ύεڞ༗ʹΑΔলϝϞϦͳܾఆ MLP(86) ଟύʔηϒτϩϯ CNN(89) ϓʔϦϯάͱΈࠐΈ ʹΑΔಛநग़ ަ੍૬ޓۭؒ๏ (06) ަߦྻʹΑΔۭؒͷؔΛߦԽ ੍૬ޓ෦ۭؒ๏ (99) ࣝผʹ༗ޮͳۭؒͷࣹӨ ಛ ந ग़ ύ λ ồ ϯ Ϛ ỽ ν ϯ ά ಛ ݕ ग़ ɾ ه ड़ ౷ ܭ త ֶ श ๏ ࠷ ۙ ୳ ࡧ SVM(95) Ϛʔδϯ࠷େԽ Random Forests(01) ΞϯαϯϒϧֶशʴϥϯμϜֶश DPM(08) Ϟσϧͷׂ latent SVM ʹΑΔࣝผ HOG(05) ޯใ Haar-like(01) box ϑΟϧλ texton(01) ϑΟϧλͷόϯΫ ਓମύʔπࣝผ 2000 2005 2010 إݕग़ ਓݕग़ ը૾ྨ εύʔεಛྔ (06) Haar-like ϐΫηϧࠩ BOF(04) ಛྔͷࣙॻԽ CHLAC(04) HLAC ʹ࣌ؒ࣠ͷՃ CARD(11) ಛྔΛ Խ AdaBoost(95) Ξϯαϯϒϧֶश αϯϓϧॏΈͷஞ࣍ߋ৽ ہॴಛྔɾ౷ܭతֶश๏ͷಈ 44**ٕज़Ϛοϓ ʮೝࣝʯͷৼΓฦΓ ϋϯυΫϥϑτಛͷਐԽ -
ը૾ೝࣝͷظͱՄೳੑɹத෦େֶֶ෦ใֶՊɹ౻٢߂ 16 %FFQ-FBSOJOHʹΑΔಛநग़ͱࣝผثͷࣗಈ֫ಘ ϙΠϯτ ɹΈࠐΈχϡʔϥϧωοτϫʔΫͷֶशʹεύʔείʔσΟϯάΛར༻ ɹ֤֊ͷग़ྗΛͯ͢౷߹͢Δํ๏ͰϩʔΧϧˍάϩʔόϧͳಛΛநग़ ಛநग़෦ ࣝผ෦ w 1FEFTUSJBO%FUFDUJPOXJUI6OTVQFSWJTFE.VMUJ4UBHF'FBUVSF-FBSOJOH<4FSNBOFU>
r ΈࠐΈ//Λ༻͍ͯਓݕग़ͷੑೳΛେ෯ʹ্ ը૾ೝࣝͷظͱՄೳੑɹத෦େֶֶ෦ใֶՊɹ౻٢߂ w 1FEFTUSJBO%FUFDUJPOXJUI6OTVQFSWJTFE.VMUJ4UBHF'FBUVSF-FBSOJOH<4FSNBOFU> r ΈࠐΈ//Λ༻͍ͯਓݕग़ͷੑೳΛେ෯ʹ্ %FFQ-FBSOJOHʹΑΔಛநग़ͱࣝผثͷࣗಈ֫ಘ ϙΠϯτ ɹΈࠐΈχϡʔϥϧωοτϫʔΫͷֶशʹεύʔείʔσΟϯάΛར༻ ɹ֤֊ͷग़ྗΛͯ͢౷߹͢Δํ๏ͰϩʔΧϧˍάϩʔόϧͳಛΛநग़ ΈࠐΈͷϑΟϧλྫ ʢ*/3*"σʔληοτɼϑΟϧλαΠζɿYʣ ݕग़ੑೳ ˠಛநग़աఔͷࣗಈ֫ಘ OS2Ͱؔ࿈ߨԋ͋Γʂ ʢମೝࣝͷ࠷લઢʣ ը૾ೝࣝͷظͱՄೳੑɹத෦େֶֶ෦ใֶՊɹ౻٢߂ w 'JOF(SBJOFE$SPXETPVSDJOHGPS'JOF(SBJOFE3FDPHOJUJPO<%FOH> r ਓ͕ͨ͠ྖҬ͔ΒಛྔΛهड़ ਓͷݟΛར༻ͨ͠ಛநग़ $SPXETPVSDJOHʹΑΓ ଟ͘ͷܦݧΛ֫ಘ ߴείΞ࣌ͷબྖҬ͔Βಛநग़ Χϥʔը૾Խ͢Δ໘ੵ͕ খ͍͞΄ͲߴείΞ ࣝผʹ༰қͳྖҬ͕ બ͞Ε͍ͯΔ ήʔϜͰߴείΞ ϙΠϯτ ɹਓ͕ࣝผʹ༰қͳྖҬΛબ͢Δ͜ͱͰࡉ͔ͳҧ͍Λࣝผ ɹ$SPXETPVSDJOHΛར༻ͯ͠େྔͷܦݧσʔλΛ֫ಘ͢Δ ήʔϜܗࣜͰਖ਼ޡఆʹ༻ͨ͠ྖҬΛબ ΫϦοΫͨ͠࠲ඪपล͕ ϒϥʔը૾͔ΒΧϥʔը૾ OS2Ͱؔ࿈ߨԋ͋Γʂ ʢମೝࣝͷ࠷લઢʣ w ۙ ࣌ʣͷಈ ը૾ೝࣝͷظͱՄೳੑɹத෦େֶֶ෦ใֶՊɹ౻٢߂ 3 ʮೝࣝʯͰऔΓѻ͏ٕज़ྖҬ ಛ ந ग़ ύ λ ồ ϯ Ϛ ỽ ν ϯ ά ಛ ݕ ग़ ɾ ه ड़ ౷ ܭ త ֶ श ๏ ࠷ ۙ ୳ ࡧ HOG(05) ޯใ Deep Learning(08) ଟχϡʔϥϧωοτϫʔΫ දݱֶश Crowdsourcing(13) ਓͷݟͷಋೖ ৄࡉը૾ࣝผ 2000 2005 2010 ਓݕग़ ը૾ྨ ਓ͕ͨ͠Ґஔ͔Βಛநग़ ϚʔΧೝࣝ BOF(04) ಛྔͷࣙॻԽ ަ੍૬ޓۭؒ๏ (06) ަߦྻʹΑΔۭؒͷؔΛߦԽ VLAD(10) ؔ࿈͢Δ VW ͷಛྔΛ༻ Fisher Vector(07) ֬ີؔʹΑΔಛྔͷදݱ ಛఆମೝࣝ Object Bank(10) ଟΫϥεͷཁૉΛಛྔԽ CoHOG(09) )0( ͷڞىදݱ εύʔεಛྔ (06) Haar-like ϐΫηϧࠩ ੍૬ޓ෦ۭؒ๏ (99) ࣝผʹ༗ޮͳۭؒͷࣹӨ Haar-like(01) box ϑΟϧλ إݕग़ ΧʔωϧτϦοΫ (00) ಛۭؒͷࣹӨ ݻ༗ۭؒ๏ (96) ࣍ݩը૾ʹΑΔ ࣍ݩମೝࣝ ૬ޓ෦ۭؒ (85) ෦ۭؒಉ࢜ͷਖ਼४֯ SIFT(99) εέʔϧෆม ಛݕग़ɾهड़ SURF(06) ੵը૾ ΞϧΰϦζϜʹΑΔߴԽ GPU SIFT(06) ϋʔυΣΞʹΑΔߴԽ FAST(06) ػցֶश ίʔφʔݕग़ BRIEF(10) ֶशແ͠ ϥϯμϜαϯϓϦϯά ORB(11) ڭࢣແֶ͠श D-BRIEF(12) ڭࢣ͋Γֶश Bin-Boost(13) ڭࢣ͋Γֶश ը૾ݕࡧ Harris-Affine(02) ΞϑΟϯෆมಛݕग़ MSER(02) ߴͳΞϑΟϯෆมಛ εϖΫτϧཧʹΑΔεέʔϧ୳ࡧ (12) ಛۭؒͷࣹӨ DAISY(08) هड़ۭؒͷվྑ HLAC(88) ߴ࣍ͷࣗݾ૬ؔ CHLAC(04) HLAC ʹ࣌ؒ࣠ͷՃ DOT(10) ޯใͷςϯϓϨʔτϚονϯά texton(01) ϑΟϧλͷόϯΫ ϚγϯϏδϣϯ ೋಛ ૿ූ߸૬ؔ (00) ًͷ૿ݮΛೋͰը૾Խ RRF(03) ํͷೱมԽ ૄςϯϓϨʔτϚονϯά (05) छྨͷϞσϧͷ͍͚ ݻ༗ղςϯϓϨʔτϚονϯά (11) ճసมԽʹؤ݈ͳใΛར༻ Co-Occurrence Template Matching(10) ݦஶੑͷߴ͍ըૉͰর߹ LBP(94) ہॴྖҬͷೋԽ CARD(11) ಛྔΛ Խ CNN(89) ϓʔϦϯάͱΈࠐΈ ʹΑΔಛநग़ SVM(95) Ϛʔδϯ࠷େԽ Random Forests(01) ΞϯαϯϒϧֶशʴϥϯμϜֶश DPM(08) Ϟσϧͷׂ latent SVM ʹΑΔࣝผ Exemplar SVM(11) ࣄྫϕʔεͷ SVM WTA Hashing(11) ߴԽ ଟΫϥεࣝผ ສΧςΰϦ ERT(06) RF ͷϥϯμϜੑΛ࠷େԽ Fern(06) RF ͷذ݅Λ֊Ͱ౷Ұ ਓମύʔπࣝผ ମݕग़ ଟΫϥε ࣗݾҐஔਪఆ Relative attribute(11) ࣮ʹΑΔؔ࿈ཁૉͷදݱ zero-shot transfer(09) ؔ࿈ཁૉ͔ΒඇֶशΫϥεͷݕग़ ηϚϯςΟοΫηάϝϯςʔγϣϯ MLP(86) ଟύʔηϒτϩϯ Online PA(06) ೖྗαϯϓϧʹԠͯ͡ॏΈϕΫτϧߋ৽ άϥεϚϯଟ༷ମ (08) ઢܗ෦ۭؒͷू߹ମ product quantization(11) αϒϕΫτϧʹΑΔྔࢠԽ Decision Jungles(13) ύεڞ༗ʹΑΔলϝϞϦͳܾఆ AdaBoost(95) Ξϯαϯϒϧֶश αϯϓϧॏΈͷஞ࣍ߋ৽ ಛྔͷࣗಈੜ ۙͷಈ ਓͷݟͷಋೖ ಛநग़ͷࣗಈԽ 44**ٕज़Ϛοϓ ʮೝࣝʯͷৼΓฦΓ $//ʹΑΔಛநग़ͱਓͷݟͷར༻ -
w ʮೝࣝʯͷޙʢ࣌ʣ 8(ʹΑΔ࣌ͷ༧ଌͱݕূ ը૾ೝࣝͷظͱՄೳੑɹத෦େֶֶ෦ใֶՊɹ౻٢߂ w ߴଟΫϥεࣝผ r 85")BTIJOHʹΑΔສΧςΰϦࣝผͷߴԽ w
ৄࡉը૾هड़ r ສΧςΰϦࣝผ ؔ࿈ཁૉʹΑΔֶश֎αϯϓϧͷϥϕϧಋग़ w [FSPTIPUMFBSOJOHʹΑΔֶश֎αϯϓϧͷదԠ r ؔ࿈ཁૉʹΑΔֶश֎αϯϓϧͷϥϕϧಋग़ సҠֶशɺܭྔֶश w େنإೝࣝ r %FFQ/FVSBM/FUXPSLͷߴԽ w ਓͱػցͷϋΠϒϦουʹΑΔೳݕࠪɺܽؕݕग़ r ਓͷݟΛಋೖͨ͠ৄࡉը૾ࣝผ w ΩʔϙΠϯτʹ͓͚Δෆมੑͷ֫ಘ r εέʔϧ୳ࡧɺΞϑΟϯมԽͷରԠ 4 ʮೝࣝʯͷޙ IUUQTXXXTMJEFTIBSFOFUTMJEFTIPXTTJJ ˚ը૾ྨଟΫϥεԽɼϞσϧେنԽ ߴԽͱͯ͠ࢬמΓɼࣝৠཹɼྔࢠԽ͕ ˚ プ ϩτλΠ プ ʹج づ ͍ͯྨ͢Δख๏ が ఏҊ͞Εɼ গγϣοτֶशʹ͓͍ͯޮՌΛൃش ˚$-*1ʹΑΓ ゼ ϩγϣοτྨՄೳͱͳͬͨ が ɼ 7-Ϟ デ ϧ で ࣮ݱ͢Δ͜ͱ༧ で ͖ͳ͔ͬͨ ˚ ビジ ϣϯج൫Ϟ デ ϧ が ಄͠େنԽͨ͠ が ɼ ߴԽͱ͍͏؍ で · だ ൃల్த ̋%-Ϟ デ ϧʹ͓͚Δ)VNBOJOUIFMPPQ が Μʹ ݚڀ͞Ε͍ͯΔ ̋&OEUPFOEֶशʹΑΔಛͷ੍ʹറΒΕͳ͍ %FUFDUJPOGSFFͳ%-ख๏ が ఏҊ ̋ɿ༧ଌ͕ਖ਼͔ͬͨ͠ɹ˚ɿҰ෦ਖ਼͔ͬͨ͠ɹºɿਖ਼͘͠ͳ͔ͬͨ ݱࡏʢ౻٢ͷࢲݟʣ -
w ʮೝࣝʯͷޙʢ࣌ʣ ༗ࣝऀʹΑΔ༧ଌͱݕূ ը૾ೝࣝͷظͱՄೳੑɹத෦େֶֶ෦ใֶՊɹ౻٢߂ Deep Learningʹද͞ΕΔਂ͍֊తߏͷֶशɾࣝผख๏ͷཧతղੳͱղੳʹجͮ͘֊ߏͷઃܭʹয͕͋ͨΓɼ͜ͷ݁ ՌΛͱʹweb্ʹଘࡏ͢Δ౷੍ͷͱΕͨҰൠతମͷը૾Ͱ͋Ε΄΅ؒҧ͍ͳ͘ೝࣝՄೳͱͳΔɽͦͷҰํͰɼΑΓ࣮ੈքدΓ ͷࡶવͱͨ͠ঢ়گʹ͓͚Δೝࣝྗ͞Εɼ·ͨɼνϟϨϯδϯάͳ՝͕ΓΛݟͤΔɽ ◦ࠓޙ͕ΓΛݟͤΔͰ͋Ζ͏՝
ɾϥΠϑϩάͷಈը૾ͷཁɼࣗવݴޠͱͷ༥߹ ΣΞϥϒϧػثͷਁಁͱڞʹɼ࣌ܥྻσʔλΛ༗ޮʹ׆༻͢Δಈը૾ͷཁٕज़͕ਐల͢Δɽ࣌ܥྻใΛ׆༻ͨ͠લޙͷจ຺ཧ ղʹΑΔೝࣝਫ਼ͷ্ͷΈͳΒͣɼਓͷײੑʹ߹கͨ͠ڵຯਂ͍γϣοτਪఆɼࣗવݴޠͰഓΘΕͨจ๏తࣝମܥ͕Ϗ δϣϯٕज़ͱ༥߹ͯ͠ैདྷͱ͞Ε͍ͯͨಈը૾ཁͷͷࢳޱͱͳΔɽ ɾίϯςϯπੜɼάϥϑΟΫεܥͱͷ༥߹ ը૾ೝࣝͱඦສϐΫηϧͷใΛ1ͭͷΧςΰϦʹԡ͠ࠐΊΔڀۃͷใѹॖٕज़ͱ͍͑ΔɽࠓޙɼάϥϑΟΫεܥͷ༥߹ ʹΑΓɼѹॖ͞Εͨใ͔Βٯʹ࣮ੈքͷใ෮ݩ͢Δίϯςϯπੜٕज़͕ਐల͢Δɽ͜ΕʹΑΓ͍จষใཧղͤͣͱਤ ΛҰຕݟΔ͜ͱʹΑͬͯॠ࣌ʹ༰ΛཧղՄೳͳใఏٕࣔज़ͷࢳޱͱͳΔɽ ɾϩϘοτϏδϣϯɼϩϘςΟΫεʢ੍ޚʣͱͷ༥߹ ౷੍ͷͱΕͨೝࣝରΛఏࣔ͢ΔͷͰ͋Εेߴ͍ࣝผੑೳ͕࣮ݱ͞ΕΔҰํͰɼೝࣝରΛ͋Β͔͡ΊఆΊͣɼࡶવͱͨ͠ը૾ ͕ೖྗঢ়گʹ͓͍ͯैདྷͷҰൠతମೝࣝख๏ͷΈͰೝࣝਫ਼͕ѱ͑͘Δٕज़ͱͯ͠΄Ͳԕ͍ɽϩϘοτͷ࣋ͭମΛ׆ ༻͢Δ͜ͱͰɼೝ͖ࣝ͢ରΛൃݟ͢Δࢹػೳͷ׆༻ͱϩϘοτͷ੍ޚٕज़ͷ༥߹ʹΑΓɼ࣮ੈքʹ͓͚ΔਅͷҙຯͰͷೳಈత ೝࣝɼֶशػೳ͕ൃల͢Δɽ ɾϓϥΠόγΞΣΞͳը૾ೝٕࣝज़ͷྲྀΕ ࠓޙΣΞϥϒϧػثͷൃల͕ݟࠐ·Ε͍ͯΔ͕ɼը૾ηϯαΛར༻ͨ͠߹ʹϓϥΠόγΛ৵͢Δը૾͕ҙਤͤͣऔಘ͞Εɼweb Ͱڞ༗͞ΕΔةݥੑΛΒΈɼը૾ηϯαΛ࣋ͭΣΞϥϒϧػثͷ֦ॆͷ͛ͱͳΔɽ͜ͷഎܠͷͱը૾ʹөΔମͷछผͷ ίϯςϯπ͕ेʹཧղՄೳͰ͋Γͳ͕ΒɼϓϥΠόγใΛͯ͢Ӆ͍͢Δٕज़͕ൃల͢ΔɽΣΞϥϒϧγεςϜͰ֫ಘͨ͠ ը૾ɼө૾ΛΫϥυιʔγϯάͳͲΛར༻ͯ͠ɼϥϕϧ༩Λߦ͏͜ͱɼwebͰͷڞ༗ɼը૾Λݟͳ͕Βͷԕִૢ࡞ͳͲʹඞ ཁෆՄܽͳٕज़ͱͳΔɽߥ͘ɼࡶવͱͨ͠ใͰՄೳͱ͢Δɽ 5 ޙͷը૾ೝࣝͷτϨϯυɿݪాୡઌੜʢ౦େʣ ը૾ೝࣝͷظͱՄೳੑɹத෦େֶֶ෦ใֶՊɹ౻٢߂ 6 ޙͷը૾ೝࣝͷτϨϯυɿ1SPG5BFLZVO,JNʢ*NQFSJBM$PMMFHF-POEPOʣ w $PNCJOFEPG3'BOE%FFQMFBSOJOH r 3BOEPN'PSFTUͱ%FFQ-FBSOJOHͷ༥߹ r ྫɿ%FDJTJPO'PSFTU<4IPUUPO> w -POHUFSNDPOUJOVPVTMFBSOJOH r OFWFSFOEJOHJNBHFMFBSOJOH r ऴΘΓͷͳ͍ը૾ֶशϑϨʔϜϫʔΫͷ࣮ݱ (a) (b) Figure 1: Motivation and notation. (a) An example use of a rooted decision DAG for classifying image patches as belonging to grass, cow or sheep classes. Using DAGs instead of trees reduces the number of nodes and can result in better generalization. For example, differently coloured patches of grass (yellow and green) are merged together into node 4, because of similar class statistics. This may encourage generalization by representing the fact that grass may appear as a mix of yellow and green. (b) Notation for a DAG, its nodes, features and branches. See text for details. input instance that reaches that node should progress through the left or right branch emanating from the node. Prediction in binary decision trees involves every input starting at the root and moving down as dictated by the split functions encountered at the split nodes. Prediction concludes when the instance reaches a leaf node, each of which contains a unique prediction. For classification trees, this prediction is a normalized histogram over class labels. Rooted binary decision DAGs. Rooted binary DAGs have a different architecture compared to decision trees and were introduced by Platt et al. [26] as a way of combining binary classifier for multi-class classification tasks. More specifically a rooted binary DAG has: (i) one root node, with in-degree 0; (ii) multiple split nodes, with in-degree 1 and out-degree 2; (iii) multiple leaf nodes, with in-degree 1 and out-degree 0. Note that in contrast to [26], if we have a C-class classification problem, here we do not necessarily expect to have C DAG leaves. In fact, the leaf nodes are not necessarily pure; And each leaf remains associated with an empirical class distribution. Classification DAGs vs classification trees. We explain the relationship between decision trees and decision DAGs using the image classification task illustrated in Fig. 1(a) as an example. We wish to classify image patches into the classes: cow, sheep or grass. A labelled set of patches is used to train a DAG. Since patches corresponding to different classes may have different average intensity, the root node may decide to split them according to this feature. Similarly, the two child nodes may decide to split the patches further based on their chromaticity. This results in grass patches with different intensity and chromaticity (bright yellow and dark green) ending up in different subtrees. However, if we detect that two such nodes are associated with similar class distributions (peaked 2ΛωοτϫʔΫঢ়ʹଓ লϝϞϦԽͱΦʔόʔϑΟοςΟϯάΛճආܾ ̋ࣗવݴޠ άϥϑΟΫεɼϩ ボ ςΟΫεͱͷ༥߹ͳͲɼݱࡏओྲྀ ͱͯ͠औΓ·Ε͍ͯΔݚڀΛత֬ʹ༧ଌ ˠಛʹɼࣗવݴޠॲཧͱͷ༥߹େ͖ͳਐలΛ げ ͨ ̋ݱࡏͷࣗݾڭࢣ͋ΓֶशܧଓֶशͷൃలΛࣔࠦ ݱࡏͷج൫Ϟ デ ϧͷߏஙʹ͓͍ͯɼॏཁͳׂΛՌ͍ͨͯ͠Δ /F3' $-*1 ࣗݾڭࢣ͋Γֶशɾܧଓֶश 6OJ"%ʢࣗಈӡసΛFOEUPFOEͰ࣮ݱʣ -
w 44**ٕज़Ϛοϓ ɿ ࣗݾڭࢣ͋Γֶशɼղ૾ɼମݕग़ɼ%ج൫Ϟσϧɼ/F3'(BVTTJBO4QMBUUJOH ϩϘςΟΫεج൫Ϟσϧɼେن ࢄֶशɼਓσʔλΛ׆༻ֶͨ͠शɼ܈ॲཧͱ%ମೝࣝɼ7JTVBM4-".ɼಈըೝࣝɾߦಈೝࣝɼΠϕϯτΧ ϝϥɼࢬמΓɼৠཹɼ%J ff
VTJPO.PEFMʗը૾ੜɼΠϝʔδϯάٕज़ɼϞσϧΞʔΩςΫνϟ 7J5 ɼϨΠΞτੜ ΫϦΤΠςΟϒϏδϣϯ ˠຊʹ͓͚Δݚڀ͕ਂ͘͞ΒʹࡉԽ ͷݱࡏ %BSL,OPXMFEHF ,OPXMFEHF%JTUJMMBUJPO <)JOUPO /*148`> ڭࢣͷ֬ʢࣝʣΛ ༻͍ͯੜెΛֶश .PEFMDPNQSFTTJPO <#VDJMV㶙 4*(,%%`> Ξϯαϯϒϧͷग़ྗΛϥϕϧͱͯ͠ ͭͷχϡʔϥϧωοτϫʔΫΛֶश ϞσϧͷΈ߹Θͤ ࣝͷछྨɾࣝͷసҠํ๏ 44**पٕज़Ϛοϓɿࣝৠཹ ෳͷڭࢣʹΑΔΞϯαϯϒϧΛར༻ .VMUJQMF5FBDIFS <:PV ,%%`> ֬Λू '&&% <1BSL,XBL &$"*`> ಛϚοϓΛू ࣗࣗͷࣝΛར༻ TFMGEJTUJMMBUJPO ਂ͍ͷࣝΛઙ͍సҠ -FBSOJOHBVOJpFEDMBTTJpFS <)PV $713`> #FZPVSPXOUFBDIFS <;IBOH *$$7`> ෳͷੜెͷΈͰֶश %.- <;IBOH $713`> ੜెؒͷࣝৠཹʹΑΓਫ਼্͕ 0/& <-BO /FVSM14`> $PMMBCPSBUJWFMFBSOJOH <4POHˍ$IBJ /FVSM14`> ੜెͷઙ͍ΛॏΈڞ༗ͯ͠ύϥϝʔλΛݮ ஈ֊తʹࣝΛసҠ 7*% <"IO $713`> ૬ޓใྔ $3% <5JBO *$-3`> ରরֶश "'% <$IVOH *$.-`> ఢରతֶश ,OPXMFEHF%J⒎VTJPO <)VBOH /FVS*14`> ֦ࢄϞσϧͷֶशํ๏ ,OPXMFEHF3FWJFX <$IFO $713`> ҟͳΔਂ͞ͷͷؒͰ ࣝΛసҠ .(% <:BOH &$$7`> ϚεΫͨ͠ੜెͷಛϚοϓ͔Β ڭࢣͷಛϚοϓΛ༧ଌ தؒͷࣝͷసҠํ๏Λվળ 3,% <1BSL $713`> αϯϓϧؒͷؔੑ 'MPXPG4PMVUJPO1SPDFEVSF <:JN $713`> ؒͷग़ྗͷ૬ޓؔ "UUFOUJPO5SBOTGFS <;BHPSVZLP *$-3`> "UUFOUJPONBQ தؒͷग़ྗ͔ΒࣝΛநग़ ".3"%*0 <3BO[JOHFS $713`> ෳͷج൫Ϟσϧ %*/0W $-*1 4". ֶशΛૣظऴྃͨ͠ڭࢣΛར༻ 3$0 <+JO *$$7`> 0OUIFF⒏DBDZ <$IPˍ)BSJIBSBO *$$7`> ೳྗΪϟοϓʹରԠ "VUP,% <-J *$$7`> தؒͷࣝදݱ &OTFNCMF,5( <0LBNPUP &$$7`> ࣝͱଛࣦͷΈ߹Θͤ ,%;FSP <-J /FVS*14`> ࣝͱଛࣦͷΈ߹Θͤ -BSHFTDBMFEJTUSJCVUFE <"OJM *$-3`> ֬Λू %VBMOFU <)PV *$$7`> ಛϚοϓΛू ෳͷੜెʹΑΔΞϯαϯϒϧΛར༻ %BUBTFU%JTUJMMBUJPO <8BOH BS9JW`> ֶशࡁΈϞσϧͷਫ਼͕ߴ͘ͳΔ Α͏ʹೖྗϊΠζΛ࠷దԽ ͦͷଞɿσʔληοτͷৠཹ ੜె ࣝΛసҠ ੜె ੜె ࣝΛసҠ ੜె ੜె ࣝΛసҠ ੜె ࣝΛసҠ ڭࢣ ڭࢣ #"/ <'VSMBOFMMP *$.-`> 4NBMMˠ4NBMMˠʜ TFMGEJTUJMMBUJPO 5FBDIFS"TTJTUBOU <.JS[BEFI """*`> -BSHFˠ.JEEMFˠ4NBMM ʢೳྗΪϟοϓʹରԠʣ %BUBEJTUPSUJPOHVJEFETFMGEJTUJMMBUJPO <9VBOE-JV """*`> ݩσʔλ͕ಉ֦͡ுޙͷσʔλͷग़ྗΛ༧ଌ ʢσʔλ͔ΒσʔλͷTFMGEJTUJMMBUJPOʣ ࣝΛసҠ ੜె σʔλ ͭͷڭࢣͰΞϯαϯϒϧ %BUBEJTUJMMBUJPO <3BEPTBWPWJD $713`> σʔλ֦ுΛར༻ 1SFQBSJOH-FTTPOT <8FO /FVSPDPNQVUJOH`> ޡೝࣝͨ͠σʔλͷࣝͱ ෆ࣮֬ͳࣝΛௐ (SBEVBM4BNQMJOH(BUF <.JOBNJ .7"`> ਖ਼ղͨ͠σʔλͷ ࣝͷΈΛసҠ ग़ྗͷࣝͷసҠํ๏Λվળ 'VODUJPO.BUDIJOH <#FZFS $713`> NJYVQʹΑΔଟ༷ͳը૾Λ༻͍ͯ ڭࢣͱੜెؒͰؔϚονϯά &⒎FDUJWFOFTTPGGVODUJPONBUDIJOH JOESJWJOHTDFOFSFDPHOJUJPO <:BTIJNB &$$78`> ϥϕϧͳ͠σʔλΛ༻͍ͯؔϚονϯά ؔϚονϯάͱͯࣝ͠ৠཹΛ࠶ߟ %*45 <)VBOH /FVS*14`> ΫϥεؒʹՃ͑ͯ Ϋϥεͷ૬ؔΛసҠ 0GGMJOF %JTUJMMBUJPO 0OMJOF %JTUJMMBUJPO ࣝΛసҠ ڭࢣ ੜె ΑΓଟ༷ͳใΛ࣋ͭ தؒͷग़ྗΛར༻ 'JU/FUT <3PNFSP *$-3`> தؒͷࣝͱͯ͠ ಛϚοϓΛ༻ ɹɹɿύϥϝʔλΛݻఆ ɹɹɿύϥϝʔλΛߋ৽ ڭࢣɿֶशࡁΈϞσϧ ੜెɿະֶशͷϞσϧ ੜెͷΈΛ༻͍ͯ ੜెؒͰࣝΛసҠ ڭࢣͷࣝΛੜెసҠ ࣝৠཹͷࣗಈઃܭ ࣝసҠΛิॿ͢ΔϞσϧΛՃ 3FTJEVBM,% <(BP BS9JW`> ࣝͷࠩΛิ͢Δ"TTJTUBOU ҟͳΔϞσϧߏؒͰࣝΛసҠ %FJ5 <5PVWSPO *$.-`> ࣝͱͯ֬͠Λ༻͍ͯ $//͔Β7J5ࣝৠཹ 0OFGPS"MM <)BP /FVS*14`> தؒग़ྗΛMPHJUۭؒʹӨ͢Δ͜ͱͰ ҟͳΔߏͷϞσϧؒͰதؒৠཹ ࣝৠཹͷࣗಈઃܭ ,5( <.JOBNJ "$$7`> ϞσϧͱଛࣦͷΈ߹Θͤ 0SBDMF,OPXMFEHF%JTUJMMBUJPO <,BOH """*`> ΞϯαϯϒϧڭࢣͷͨΊͷੜెͷϞσϧߏ ΫϥεߏλεΫ͕ҟͳΔෳͷڭࢣͷࣝΛੜెʹू 4UVEFOUCFDPNJOHUIFNBTUFS <:F $713`> ηϚηάΛֶशͨ͠ڭࢣͱਂਪఆΛֶशͨ͠ڭࢣ "NBMHBNBUJOH,OPXMFEHF <4IFO """*`> ҟͳΔྨλεΫΛֶशͨ͠ෳͷڭࢣ ಛఆͷλεΫ ֶश Ϟσϧʹ͓͚ΔࣝΛઃܭ $-*1,% <'BOH $713`> $-*1ɿ$-*1ʹ͓͍ͯ ैདྷͷࣝͷ༗ޮੑΛௐࠪ .JOJ7J5 <;IBOH $713`> 7JTJPO5SBOTGPSNFSɿ ΞςϯγϣϯॏΈͱύοντʔΫϯ .BOJGPME%JTUJMMBUJPO <)BP /FVS*14`> 7JTJPO5SBOTGPSNFSɿ ύονؒͷؔੑ -BSHFTDBMFJODSFNFOUBMMFBSOJOH <8V $713`> ܧଓֶशɿաڈλεΫͰ ֶशͨ͠Ϟσϧͷ֬ *NQSPWJOHGBTUTFHNFOUBUJPO XJUIUFBDIFSTUVEFOUMFBSOJOH <9JF #.7$`> ηϚηάɿۙͷϐΫηϧͱͷMPHJUؔ 4&&% <'BOH *$-3`> ࣗݾڭࢣ͋Γֶशɿ αϯϓϧؒͷؔੑ -FBSOJOHF⒏DJFOUPCKFDUEFUFDUJPO NPEFMTXJUILOPXMFEHFEJTUJMMBUJPO <;BHPSVZLP *$-3`> ମݕग़ɿମྖҬͷۣܗ ڭࢣ ੜె ࣝΛసҠ ੜె ੜె ࣝΛసҠ ੜె ੜె ࣝΛసҠ ࢬמΓͷ ϋʔυΣΞରԠ &*&<4)BO *4$"> ॏΈڞ༗ͱૄߦྻԋࢉʹରԠ /7*%*""NQFSF</7*%*" (5$> ࢬמΓޙͷਪΛߴԽ͢ΔػೳΛ࣮ %31"*<3FOFTBT *44$$`> ϚΠΫϩϓϩηοαͰ͋Δ3;7)ʹࡌ͞ΕΔ"*ΞΫηϥϨʔλ ޯϕʔε Ϛάχνϡʔυϕʔε 5SBOTGPSNFS $// ๅ͘͡Ծઆ 44**पٕज़ϚοϓɿࢬמΓ 5SBOTGPSNFSϞσϧͷొ "UUFOUJPO*T"MM:PV/FFE <7BTXBOJ /*14`> $//Ϟσϧͷొ -F/FU <-F$VO > ࢬמΓͷ࢝·Γ $//Ϟσϧͷൃల "MFY/FU <,SJ[IFWTLZ /*14`> આ໌ੑΛߟྀͨ͠ख๏ ϨΠϠʔ୯ҐͰࢬמΓ ࢬמΓޙͷඍௐෆཁ ଞͷѹॖख๏ͱͷซ༻ ֶशલʹࢬמΓՄೳ ߏԽࢬמΓ 4USVDUVSFE1SVOJOH ඇߏԽࢬמΓ 6OTUSVDUVSFE1SVOJOH ࢉग़ͨ͠ޯΛجʹͯ͠ࢬמΓ ॏΈύϥϝʔλͷେ͖͞Λجʹͯ͠ࢬמΓ ߏԽࢬמΓ 4FNJTUSVDUVSFE1SVOJOH ɿֶशલʢॳظঢ়ଶͷϞσϧͷॏΈΛධՁʣ ɿֶशதʢεύʔεਖ਼ଇԽͰϞσϧͱϚεΫΛֶशʣ ɿֶशޙʢֶशࡁΈϞσϧͷॏΈΛධՁʣ ɿਪ࣌ʢೖྗʹԠͯ͡༻͢ΔॏΈΛಈతʹબʣ (BUF%FDPSBUPS <:PV /*14`> ςΠϥʔల։ϕʔεͷ εέʔϦϯάϑΝΫλͷධՁ 'JMUFS1SVOJOH WJB(FPNFUSJD.FEJBO <)F $713`> ΧʔωϧͷॏΈٴͼ Β͖ͭͷେ͖͞Λߟྀ 3FUIJOLJOHUIF4NBMMFS/PSN -FTT*OGPSNBUJWF"TTVNQUJPO <:F *$-3`> ਖ਼ଇԽ߲ʹܭࢉྔͷج४Λಋೖ 3FT3FQ <%JOH *$$7`> ͳॏΈʹͷΈਖ਼ଇԽΛద༻ *NQPSUBODF&TUJNBUJPO <.PMDIBOPW $713`> ΧʔωϧΛআͨ͠ࡍͷ ଛࣦͷมԽྔͷೋΛධՁ (SPVQ'JTIFS1SVOJOH <-JV *$.-`> 'JTIFSJOGPSNBUJPOΛ ༻͍ͨॏΈͷධՁ $1SVOF <,JN &$$7`> %//ίϯύΠϥΛ༻͍ͨ ࣮ߦϕʔεͷࢬמΓ /FU"EBQU <:BOH &$$7`> ࣮ߦ࣌ؒΛج४ʹධՁ /FUXPSL4MJNNJOH <-JV *$$7`> εέʔϦϯάϑΝΫλͷ ಋೖͱਖ਼ଇԽ $-*12 <5VOH $713`> ϚάχνϡʔυʹΑΔࢬמΓͱྔࢠԽΛซ༻ -FBSOJOH#PUI 8FJHIUTBOE$POOFDUJPOT GPS&GGJDJFOU/FVSBM/FUXPSLT <)BO "$.`> ॏΈͷେ͖͞Λج४ʹࢬמΓ %ZOBNJD/FUXPSL4VSHFSZGPS &GGJDJFOU%//T <(VP /*14`> ࢬמΓ͞ΕͨॏΈΛ ճ෮ͤ͞ΔͨΊͷΦϓγϣϯΛՃ &YQMPSJOH4QBSTJUZJO 3FDVSSFOU/FVSBM/FUXPSLT </BSBOH *$-3`> 3//ͷֶशաఔͰޮతʹద༻ 1SVOJOH%FFQ/FVSBM/FUXPSLT GSPNB4QBSTJUZ1FSTQFDUJWF <%JBP *$-3`> ܦࡁֶͷͷ๏ଇΛ༻͍ͨࢬמΓͷಈతܾఆ -FBSOBCMF1SVOJOH <:BP BS9JW`> ༧Ίઃఆͨ͠ࢬמΓʹରԠͨ͠ਖ਼ଇԽख๏ 4/*1 <-FF *$-3`> ଛࣦؔʹର͢ΔॏΈͷ ײΛධՁ 4ZOGMPX <5BOBLB /*14`> ߴѹॖʹΑΔͷ่յରԠ ʢσʔλϑϦʔʣ 0QUJNBM#SBJO4VSHFPO <)BTTJCJ4UPSL /*14`> ͋ΔॏΈ̍ͭʹର͢ΔςΠϥʔల։ Λ༻͍ͯ0#%ΛҰൠԽ (SB41 <8BOH *$-3`> γΞϯΛ༻͍ͯ ॏΈಉ࢜ͷؔΛධՁ /FVSBM5BOHFOU5SBOTGFS <-JV *$.-`> /5,Ͱಛֶ͚ͮͨशμΠφϛΫε Λ฿͢ΔαϒωοτϫʔΫΛ୳ࡧ 0QUJNBM#SBJO%BNBHF <-F$VO /*14`> γΞϯΛ༻͍ͯ ଛࣦ͕૿Ճ͠ͳ͍ॏΈΛಛఆ 5IF-PUUFSZ5JDLFU)ZQPUIFTJT <'SBOLF *$-3`> ॳظωοτʹ ಉੑೳͷαϒωοτ͕ଘࡏ 4USPOH-PUUFSZ5JDLFU)ZQPUIFTJT <3BNBOVKBO $713`> େنͳωοτʹະֶशͰ ಉੑೳͷαϒωοτ͕ଘࡏ %VBM-PUUFSZ5JDLFU )ZQPUIFTJT <#BJ *$-3`> ϥϯμϜʹબͨ͠αϒωοτ ͨΓ͘͡ʹม͑Δ͜ͱ͕Մೳ 1BUDI4MJNNJOH <5BOH $713`> ύονؒͷྨࣅΛ؍ "OBMZ[JOH .VMUJ)FBE4FMG"UUFOUJPO <7PJUBM "$-`> -ਖ਼ଇԽΛ༻͍ͯϔουΛআ 4USVDUVSFE1SVOJOHPG -BSHF-BOHVBHF.PEFMT <8BOH "$-`> ର֯ߦྻ)BSE$PODSFUF ͔ΒαϯϓϦϯά 7JTJPO5SBOTGPSNFS1SVOJOH <;IV ,%%`> ֤ͷલޙͰτʔΫϯͷ ࣍ݩΛѹॖ͢ΔͨΊͷ ର֯ߦྻΛಋೖ 6OJGJFE7JTVBM5SBOTGPSNFS $PNQSFTTJPO <:V *$-3`> ࢬמΓ ϒϩοΫεΩοϓ ࣝৠཹ 91SVOFS <:V *$$7`> 7J5ͷઆ໌ੑΛߟྀͨ͠ࢬמΓ 8%1SVOJOH <:V """*`> ͷਂ͞ͱ෯Λಉ࣌ʹݮ $IBOOFM1SVOJOH <)F *$$7`> ͷલޙͰಛྔͷ มԽ͕࠷খʹͳΔ Α͏ͳॏΈΛ࠶ߏ %JTDSJNJOBUJPOBXBSF $IBOOFM1SVOJOH <;IVBOH /*14`> ຖʹࣝผثΛಋೖ %ZOB#&35 <)PV /*14`> ࢬמΓ ৠཹͰͱ෯Λѹॖ 304*5" <-JV """*`> ࢬמΓʴϥϯΫҼղ ʴࣝৠཹͷซ༻ 050 <$IFO /*14`> ύϥϝʔλΛάϧʔϓԽͯ͠ άϧʔϓ୯ҐͰࢬמΓ "50 <$IFO $713`> ίϯτϩʔϥωοτϫʔΫ ͰϚεΫΛֶश ;FSP51SVOF <8BOH $713`> τʔΫϯͷॏཁͱ ྨࣅͰࢬמΓ *UFSBUJWF4/*1 <%F+PSHF *$-3`> 4/*1ΛϚϧνγϣοτͰద༻ /5,4"1 <8BOH *$-3`> /5,ͷมԽ͕খ͍͞ॏΈΛධՁ %FFQ.P& <8BOH 1.-3`> ֶशͨ͠ήʔτωοτϫʔΫ ͰಈతʹνϟϯωϧΛࢬמΓ %FKB7V <-JV *$.-`> ೖྗʹԠͯ͡ಈతʹ νϟωϧϔουΛΓସ͑ +BNCB <-JFCFS BS9JW> .P&తʹαϒωοτ ϫʔΫΛΓସ͑ 1SVOJOH'JMUFSJO'JMUFS <.FOH /*148`> ΈࠐΈϑΟϧλΛ ετϥΠϓ୯ҐͰࢬמΓ /1"4 <-J $713`> ҟͳΔΧʔωϧαΠζʹదԠͨ͠ ίϯύΠϥϕʔεͷࢬמΓ 1$0/7 <.B """*`> &-P(ʹΑΔϑΟϧλͷ ύλʔϯࢬמΓͱίϯύΠϥ 4QBSTF(15 <'SBOUBS *$.-`> ϒϩοΫαΠζͷࢦఆʹΑΓ ߏతͳεύʔεੑʹ֦ுՄೳ -FBSOJOH4USVDUVSFE4QBSTJUZ JO%FFQ-FBSOJOH <8FO /*14`> Χʔωϧɼνϟϯωϧɼͷਂ͞ ʹରͯ͠ਖ਼ଇԽΛద༻ &BSMZ#JSEUJDLFUT <:PV *$-3`> ૣֶ͍शஈ֊ͰͨΓ͘͡Λൃݟ "SF4JYUFFO)FBET 3FBMMZ#FUUFSUIBO0OF <.JDIFM /*14`> ϔουຖͷޯͷେ͖͞Λج४ʹࢬמΓ $IBTJOHTQBSTJUZ JOWJTJPOUSBOTGPSNFST <$IFO /*14`> ϔουͷग़ྗಛΛςΠϥʔల։ ΛԠ༻ͯ͠ධՁ 4JOHMFTIPU1SVOJOH 'PS1SFUSBJOFE.PEFMT <,PIBNB *$$78`> ࣄલֶशࡁΈϞσϧʹରԠ %FQ(SBQI <'BOH $713`> %//ͷ֤ͷґଘؔΛ άϥϑͱͯ͠දݱ ࢬמΓͷλΠϛϯά 8JOOJOH-PUUFSZ5JDLFUT JO%FFQ(FOFSBUJWF.PEFMT <,BMJCIBU """*> ੜϞσϧʹ֦ு (SBQI-PUUFSZ5JDLFU (-5 )ZQPUIFTJT <$IFO *$.-> (SBQI/FVSBM/FUXPSLʹ֦ு 5PXBSET4USVDUVSBMMZ4QBSTF-PUUFSZ5JDLFUT <$IFO *$.-`> ඇθϩཁૉΛάϧʔϓԽͯ͠ ߏతʹεύʔεͳͨΓ͘͡Λ֫ಘ ʢத෦େ.13(ʴ%FOTP*5-BCͰ࡞ʣ -
w ج൫Ϟσϧʢ'PVOEBUJPONPEFM େྔ͔ͭൣͳσʔλͰֶश༷ͯ͠ʑͳԼྲྀλεΫʹసҠͰ͖ΔਂֶशϞσϧ ࣌ʹ༧ଌͰ͖ͣɼݱࡏओྲྀͱͳٕͬͨज़ͷҰͭ FH4FHNFOU"OZUIJOH.PEFM 4".
<,JSJMMPW *$$7> %*/0W<0RVBC BS9JW> w ج൫ϞσϧͷϚϧνϞʔμϧԽ $-*1 $POUSBTUJWF-BOHVBHF*NBHF1SFUSBJOJOH <3BEGPSE *$.-> ;FSPTIPUΫϥεྨͱPQFOWPDBCVMBSZೝࣝͷల։ ͷݱࡏ I1·T2 I1·T3 … I2·T1 I2·T3 … I3·T1 I3·T2 … ⋮ ⋮ ⋮ I1·T1 I2·T2 I3·T3 (1) Contrastive pre-training Image Encoder Text Encoder Pepper the aussie pup Pepper the aussie pup Pepper the aussie pup Pepper the aussie pup T1 T2 T3 … I1 I2 I3 ⋮ (2) Create dataset classifier from label text plane car dog ⋮ bird A photo of a {object}. ⋮ Text Encoder T1 T2 T3 TN … (3) Use for zero-shot prediction Image Encoder I1 I1·T2 I1·TN I1·T1 … … A photo of a dog. TN IN·T1 IN·T2 IN·T3 I1·TN I2·TN I3·TN ⋮ … IN … ⋮ ⋱ IN·TN I1·T3 w $*'"3ͷը૾ͱϓϩϯϓτͷಛྔΛ6."1Ͱ࣍ݩʹ࣍ݩݮͯ͠ՄࢹԽ ͳͥ;FSPTIPUJNBHFDMBTTJpDBUJPO͕Մೳʁ ɿը૾ͷಛྔ ɿϓϩϯϓτͷಛྔ Ϋϥε͝ͱʹಛྔ͕ಠཱ͠ɼը૾ͱϓϩϯϓτ͕ಉ͡ՕॴʹຒΊࠐ·Ε͍ͯΔ BQIPUPPGBBJSQMBOF BQIPUPPGBCJSE BQIPUPPGBDBU BQIPUPPGBEFFS BQIPUPPGBGSPH BQIPUPPGBIPSTF BQIPUPPGBBVUPNPCJMF BQIPUPPGBEPH BQIPUPPGBTIJQ BQIPUPPGBUSVDL Query: saturn V, blossom Query: golden gate, yacht Query: Oculus, Ukulele Figure 5. Open-vocabulary segmentation with user-defined queries. Our model accurately segments unseen categories, such as the Saturn V rocket, Oculus headset, and Golden gate bridge. 150 and PC-59 We hypothesis that 12K nouns are adequate for the CLIP to retain its open-vocabulary ability. Thus, we choose to use 1 caption for efficiency purposes as it’s 5x faster in training then using 5 captions. 4.3.2 Mask prompt tuning We ablate the effect of mask prompt tuning in Table 3. The baseline model is MaskFormer Swin-Base with CLIP ViT-L/14. If we only use mask prompt tuning (case (a)), our model outperforms the baseline by a large +4.7% and +4.0% mIoU improvement on ADE-150 and PC-59, respec- tively. Case (b) shows the result of full model fine-tuning. Although it achieves the best accuracy, the trainable param- eters are orders of magnitude higher. In contrast, the pro- posed mask prompt tuning only modifies the input without changing CLIP’s weight. Furthermore, mask prompt tun- ing can further improve over a fully finetuned model, as shown in case (c). Case (c) achieves 29.6% mIoU ADE- GT: building Pred: skycraper GT: rail Pred: road Figure 6. Ambiguity of the class definition in open vocabulary segmentation evaluation. fact that language defined categories are ambiguous and can overlap with each other. Designing a better evaluation met- $-*1ʹΑΔରরֶश ը૾ಛͱݴޠͷରԠ͚ͮ 0QFOWPDBCVMBSZೝࣝ 074FH<'-JBOH $713> -
w 44**ٕज़Ϛοϓ Λ࡞ͨ͜͠ͱ͕͖͔͚ͬͱͳΓऔΓΜͩݚڀ զʑ .13( ͷऔΓΈ ͚͠ ࠷ऴత
Γɼೝ BN Λߟྀ ͕ײ Δɽ͜ खͰम νϡʔ map Α͏ʹ ͚Δਓ ޮੑ ਓखʹ ɼग़ྗ Δͨ tention map ͷେ͖͞ 14×14 Ͱ͋Δɽ࣍ʹɼ Attention map Λमਖ਼͢ΔͨΊʹɼAttention map Λ 224×224 ʹ ֦େ͠ɼAttention map ΛਓखʹΑΓमਖ਼͢Δɽਤ 2 ͷྫͷ߹ɼਖ਼ղϥϕϧ͕ “Dalmatian” Ͱ͋Δը૾Λ ResNet-152+ABN ೖྗͨ͠ͱ͖ɼೋͭͷମ͕ը૾ தʹؚ·Ε͍ͯΔͨΊ “Soccer ball” ͱޡೝࣝͨ͠ɽ͜ ͷͱ͖ɼAttention map ΛՄࢹԽ͢Δͱɼ“Soccer ball” ʹରͯ͠ڧ͘ࢹ͍ͯ͠Δ͜ͱ͕֬ೝͰ͖Δɽͦ͜Ͱɼ ਤ 2 Attention map ͷमਖ਼ํ๏ 𝑚 3 𝑚 1 𝑚 2 𝐿^ 𝑦,1 𝐿^ 𝑦,2 𝐿^ 𝑦,3 𝐿 1,2 𝐿 1,3 𝐿 2,1 𝐿 3,1 𝐿 3,2 𝐿 2,3 ^ 𝑦 ^ 𝑦 ^ 𝑦 ήʔτؔ • 5ISPVHI(BUF • $VUPGG(BUF • -JOFBS(BUF • $PSSFDU(BUF • 3FT/FU ධՁରϊʔυ • 3FT/FU • 3FT/FU • 8JEF3FT/FU ิॿϊʔυ ϋΠύϥ୳ࡧ ਂֶशϞ デ ϧͷਓͷݟͷΈࠐΈ ୳ࡧʹΑΔൃݟతݚڀΞ プ ϩʔν "UUFOUJPO#SBODI/FUXPSL <'VLVJ $713> <ࡾݪ ৴ֶ> ΞςϯγϣϯϚοϓΛਓͷݟʹΑΓखಈͰ"#/ϑΝΠϯνϡʔχϯά ˠ デ ʔλ ド Ϧ ブ ϯͳ"*Ϟ デ ϧͱਓͷϋΠ ブ Ϧο ド ʹΑΔΞ プ ϩʔν ݚڀऀ が ઃܭ͍ͯͨͨ͠Ίݶఆత で ͋ͬͨෳϞσϧؒͷࣝৠཹΛϋΠ ύϥ୳ࡧͯ͠࠷దԽˠ৽ͨͳࣝసҠ๏ が ൃݟ͞Εͨ ,OPXMFEHF5SBOTGFS(SBQI <.JOBNJ "$$7> <0LBNPUP &$$7> -
ૉਓൃݰਓ࣮ߦ த෦େֶϩΰ த෦େֶϩΰ w ࣝసҠάϥϑΛఏҊࣝసҠ グ ϥϑʹΑΔڞಉֶश<.JOBNJ .*36 "$$7>
ݚڀऀઃఆͱղ͖ํΛݶఆͤ ず ɼࣝసҠΛදݱ͢ΔϑϨʔϜϫʔΫΛݚڀऀ が ઐੑΛൃ شͯ͠ઃܭ ࠷దԽ୳ࡧΛߦ͏͜ͱ で ৽ͨͳݟΛൃݟ 4PGUXBSF ࣌ͷݚڀํ๏ ܭࢉػ が େͳ パ ϥϝʔλۭ͔ؒΒ༷Λຬͨ͢ プ ϩ グ ϥϜΛ୳ࡧ ʮૉਓൃݰਓ࣮ߦʯ ୳ࡧʹΑΓ֫ಘͨ͠৽ͨͳݟ が ࣍ͷ৽ͨͳݚڀͷॹͱͳΔ͜ͱΛظ ݚڀۭؒ ͜Ε·Ͱͷݚڀ ݚ ڀ ͷ ෳ ࡶ ͞ ૉਓൃݰਓ࣮ߦ -
w 5BFLZVO,JNઌੜ ,"*45 ʹ࠶ଧʢ݄ʣ ະདྷʹ͚ͯ Hi! It's great to
hear from you. Yes, the continuous learning and self-supervised learning approaches are indeed pivotal for the development of AGI. Google's Pathways architecture is particularly fascinating, as it aims to leverage a mixture of experts (MoE) strategy by training numerous neural networks on a variety of tasks and selectively activating a few for each inference. This method, while computationally intensive, represents a promising step towards AGI by efficiently managing computational resources and creating more specialized pathways for different inputs. The idea of combining a binary decision tree with deep neural networks to form a hierarchical MoE is also intriguing. It blends the interpretability of random forests with the powerful learning capabilities of deep neural networks, potentially offering a more transparent yet highly effective model. This hybrid approach could address some of the black-box issues inherent in deep learning models while maintaining high performance across diverse tasks. It’s an exciting time for AI research, and these advancements are paving the way for more sophisticated and capable AI systems. ͦ͏Ͱ͢Ͷɺܧଓֶशࣗݾڭࢣ͋ΓֶशͷΞϓϩʔνɺAGIʢਓ൚༻ೳʣͷ։ൃ ʹ͓͍ͯඇৗʹॏཁͰ͢ɻGoogleͷPathwaysΞʔΩςΫνϟಛʹڵຯਂ͍Ͱ͢ɻ͜Ε ɺΤΩεύʔτͷࠞ߹ʢMoEʣઓུΛ׆༻ͯ͠ɺ͞·͟·ͳλεΫʹରͯ͠ଟͷχϡ ʔϥϧωοτϫʔΫΛ܇࿅͠ɺਪ͝ͱʹબతʹ͍͔ͭ͘Λ׆ੑԽ͢Δ͜ͱΛࢦͯ͠ ͍·͢ɻ͜ͷํ๏ܭࢉෛՙ͕ߴ͍ͷͷɺܭࢉϦιʔεΛޮతʹཧ͠ɺҟͳΔೖྗ ʹରͯ͠ΑΓઐԽ͞Εͨܦ࿏Λ࡞͢Δ͜ͱͰɺAGIʹ͚ͨ༗ͳҰาͱͳΓ·͢ɻ όΠφϦܾఆͱਂχϡʔϥϧωοτϫʔΫΛΈ߹Θͤͯ֊తͳMoEΛܗ͢Δͱ ͍͏ΞΠσΞڵຯਂ͍Ͱ͢ɻ͜ΕɺϥϯμϜϑΥϨετͷղऍՄೳੑͱਂχϡʔϥ ϧωοτϫʔΫͷڧྗͳֶशೳྗΛ༥߹ͤ͞ɺΑΓಁ໌ੑ͕͋Γͳ͕ΒߴੑೳͳϞσϧΛ ఏڙ͢ΔՄೳੑ͕͋Γ·͢ɻ͜ͷϋΠϒϦουΞϓϩʔνɺਂֶशϞσϧʹࡏ͢Δ ϒϥοΫϘοΫεΛղܾ͠ͳ͕Βɺଟ༷ͳλεΫͰߴ͍ੑೳΛҡ࣋Ͱ͖Δ͔͠Ε· ͤΜɻ AIݚڀʹͱͬͯඇৗʹΤΩαΠςΟϯάͳ࣌ظͰ͋Γɺ͜ΕΒͷਐาɺΑΓߴͰೳྗ ͷߴ͍AIγεςϜͷಓΛΓ։͍͍ͯ·͢ɻ (PPHMF1BUIXBZTIUUQTXXXZPVUVCFDPNXBUDI W /GE$D&;X "(*ʹ͚ͨ.P& .JYUVSFPG&YQFSUT ઓུ ෳͷઐՈϞσϧΛΈ߹ΘͤͯશମͷੑೳΛ্ͤ͞ΔΞϓϩʔν ޮతͳܭࢉɿ֤ೖྗʹରͯ͠Ұ෦ͷઐՈϞσϧͷΈ͕׆ੑԽ͢ΔͨΊɼɹ େنͳϞσϧʹൺͯܭࢉίετ͕͍ ൚༻ੑͷ্ɿҟͳΔઐՈ͕ҟͳΔλεΫʹಛԽ͢ΔͨΊɼશମͱͯ͠ Ϟσϧ͕෯͍λεΫʹରԠՄೳ ҟͳΔͳΔϞʔμϦςΟʢςΩετ ը૾ ԻͳͲʣΛ౷߹తʹֶश͢Δ ͜ͱͰɼҟͳΔσʔλιʔε͔ΒͷใΛΈ߹ΘͤͯΑΓแׅతͳཧղ ΛಘΔ͜ͱ͕Մೳ -
w 44**ͷٕज़Ϛοϓ࡞ΛৼΓฦΔͱʢݸਓతײʣ 👎اըͨ͠ྑ͍͕େมͳ࡞ۀ 👎ϚοϓʹهࡌͰ͖ͳ͍ଞͷٕज़ଟ͋Γ 👌ԿΛ͖͔͢ͷඪΛݟग़ͤͨ w ҰਓͰٕज़ಈΛѲ͢Δ͜ͱ͕͍࣌͠ʹ 44**DWQBQFSDIBMMFOHFͷ৫Λ͑ͨݚڀऀɾ։ൃऀίϛϡχςΟͷ׆༻ Χϝʹ;͞Θֶ͍͠ͼͷڥˠ44**ίϛϡχςΟ
J ͖ͳՊझຯ͕ͱ͜ͱΜٻͰ͖ɼ JJ ͦͷ͜ͱ͕पΓ͔Βଚܟ͞Εɼ JJJ ͦͷಛٕΛؒʹڭ͑Δ͜ͱͰपΓʹՃ͑ͯ ຊਓ·Ͱֶ͕ͼΛਂΊΒΕΔڥ ҏ౻ެฏʰେֶͱຊͷةػ － ࠶ߟʱIUUQTXXXLFJPBDKQKBBCPVUQSFTJEFOUCMPH 44**ٕज़ϚοϓΛ௨ͯ͠ݟΔաڈɾݱࡏɼͦͯ͠ະདྷ ޙͷ44**ʹ͓͍ͯ44**ٕज़ϚοϓΛҰॹʹ࡞͠·͠ΐ͏ʂ -
44**ٕज़Ϛοϓ ࡞ڠྗऀ ࢁޱम ౦ࣳ ࢁԼོٛ த෦େֶ
ঙҳ ిؾ௨৴େֶ Ҫਓ αϜεϯຊݚڀॴ ്ܓೋ ۀɾ৯࢈ۀٕज़૯߹ݚڀػߏ ਖ਼थʢιχʔʣ Ҫ৲ળٱ ΦϜϩϯ ʢॱෆಉɼॴଐ࣌ʣ औΓ·ͱΊɿ ਡ๚ਖ਼थʢΦϜϩϯʣ Ӝ৻࡞ʢࡕେʣ ౻٢߂ʢத෦େֶʣ ௗډळʢ౦ژۀେֶʣ ڠྗऀɿ ҆ഒຬ σϯιʔΞΠςΟϥϘϥτϦ ೖߐ߽ /55 ؠଜխʔ େࡕཱେֶ 5BF,ZVO,JN *NQFSJBM$PMMFHF-POEPO খྛঊ ࢈ۀٕज़૯߹ݚڀॴ ࡔ্จ ໊ݹۀେֶ ᓎా݈ ౦େֶ ౡాರ࢜ भେֶ ਿຊໜथ ౦ژۀେֶ ۄప ౡେֶ ୩ઍՃ ࢈ࣗಈं ݪҰ भେֶ தࢁӳथ ౦ژେֶ ݪాୡ ౦ژେֶ ງాೋ ౦ژେֶ ߁ത ಸྑઌՊֶٕज़େֶӃେֶ
hf149
institutdelelevage
kzmssk
taniai
izumiyama_lab
pman0214
kanojikajino
resolutionfoundation
ipullrank
chriscoyier
szymonslowik
soysaucechin
mfonobong
mraible
jacobtomlinson
jct
addyosmani
tammyeverts
coloredviolet
popppiees
