DSCC PARALLEL DEEP LEARNING ENSEMBLES FOR HUMAN POSE ESTIMATION

Transcription

1 Proceedgs of the ASME 08 Dyamc Systems ad otrol oferece DS08 Setember 30-October 3, 08, Atlata, Georga, USA DS PARALLEL DEEP LEARNING ENSEMBLES FOR HUMAN POSE ESTIMATION Hal Re Det. of Mechacal Egeerg Vrga Tech, Al Kumar Det. of Mechacal Egeerg Vrga Tech, Xra Wag Det. of Mechacal Egeerg Vrga Tech, Phas Be-Tzv* Det. of Mechacal Egeerg Vrga Tech, ABSTRAT Ths aer resets a effcet method to detect huma ose wth moocular color magery usg a arallel archtecture based o dee eural etwork. The etwork reseted ths aroach cossts of two sequetally coected stages of 3 arallel NN esembles, where each esemble s traed to detect oe secfc kd of lkage of the huma skeleto structure. After detectg all skeleto lkages, a votg scorebased ost-rocessg algorthm assembles the dvdual lkages to form a comlete huma structure. Ths algorthm exlots huma structural heurstcs whle assemblg skeleto lks ad searches oly for adjacet lk ars aroud the exected commo jot area. The use of structural heurstcs the reseted aroach heavly smlfes the ost-rocessg comutatos. Furthermore, the arallel archtecture of the reseted etwork eables mutually deedet comutg odes to be effcetly deloyed o arallel comutg devces such as GPUs for comutatoally effcet trag. The roosed etwork has bee traed ad tested o the OO 07 erso-keyots dataset ad delvers ose estmato erformace matchg state-of-art etworks. The arallel esembles archtecture mroves ts adatablty alcatos amed at detfyg oly secfc body arts whle savg comutatoal resources. Keywords: Pose Estmato, ovolutoal Neural Networks (NN), Lkage-based Aroach, Parallel NN Archtecture. INTRODUTION As recet rogress comutatoal caabltes eable maches to come to the real world from a lab settg, t becomes mortat to uderstad ad study earby huma actvty to address safety cocers. Huma ose estmato s already a actve research roblem for mache erceto *orresodg author betzv@vt.edu systems self-drvg cars, search ad rescue systems, automated survellace ad other Huma-Robot Iteracto (HRI) alcatos []. Accurate ad effcet huma ose estmato s crtcal achevg hgh-level tasks such as edestra avodace, automated robotc lftg ad movg vctms for search ad rescue alcatos, huma behavor recogto, etc. Based o the alcato ad the avalablty of sesg modalty, the ut data could be ether D mages [ 6], 3D ot clouds [7,8], oe sgle frame or a sequece of frames (moto-trackg) [9]. I recet decades, researchers have focused o model-based algorthms, whch deloy fely tued feature extractors such as SIFT [5] ad HoG [0] alog wth dfferet huma models such as Pctoral Structures [] ad Actve Shae Models [6]. I more recet years, as artfcal tellgece has become sgfcatly oular wth the HRI researchers due to the advacemets comutg techologes, exlorato of eural etworks for huma ose estmato has also cked u the ace. O the same track, Toshev ad Szegedy [] utlzed sequetally coected covoluto layers ad fully coected layers to buld oe dee eural etwork for hgh recso ose estmato. L et al. roosed a ose-jot reressor wth body-art detector usg a sgle dee eural etwork [4]. I a slghtly dfferet aroach, Tomso et al. roosed a hybrd archtecture usg both covolutoal eural etworks ad Markov Radom Feld [3]. Whle other researchers exlored huma ose estmato stll-magery, the comuter vso grou from U Berkeley looked to the use of Recurret-NNs for both ose estmato ad gat/acto recogto vdeo ut [,3]. Recogto of the ose of a sgle erso [5,6,4 8] a mage sets u the foudato for ose estmato of multle ersos [9,,9 ]. Oe straghtforward aroach for mult-erso oseestmato s to aly a erso detector o the ut mage, ad the for each erso detector roosal, aly a sgle erso oyrght 08 ASME Dowloaded From: htts://roceedgs.asmedgtalcollecto.asme.org o /3/08 Terms of Use: htt://

2 ose estmato method. Ths aroach s called a to-dow method [,] ad suffers from early commtmet, meag that there s o chace to detect a erso that s ot roosed by the erso detector, such as Faster RNN [3]. The comutatoal cost of ths aroach s roortoal to the umber of detector roosals from oe mage. Aother strategy for mult-erso ose estmato s the bottom-u method [9]. I these models, the etwork detects the body arts of the erso vsble the ut magery frst ad the assembles them to multle dvduals accordg to a gve olcy. Ths kd of aroach solves the early commtmet roblem but suffers from a comutatoal cost roblem, sce assocatg dfferet arts to dvduals s a NP-hard roblem [0]. Oe such examle of the bottom-u aroach bulds uo a dee covolutoal etwork to detect jots. A hgher-level satal model s the used to costra jot ter-coectvty ad geerate the global ose [3]. Isafutdov et al. used a strog detector to detect erso's jots ad assemble those jots usg mage-codtoed arwse terms [9]. Oe hybrd aroach cororatg both the bottom-u ad to-dow methods has bee develoed. Sheg et al. roosed a method by whch oe bottom-u detector s used to detect jots ad oe to-dow huma detector s used to rule out the bottom-u false alarms resultg sgfcatly mroved trackg accuracy [4]. ao et al. roosed the use of a jot heat ma wth art affty felds to estmate D huma oses by usg mult-stage, sequetally coected NN braches [5]. I ther aroach, each brach was resosble for redctg oe secfc body lkage terms of jot heat mas ad a D vector feld (Part Affty Felds (PAF)) reresetg the drecto shae of the lkage mask. Ths aer bulds o the work by ao et al. [5] ad resets a dee eural etwork archtecture for mult-erso ose estmato usg arallel NN esembles. Smlar to revous work, the roosed eural etwork assumes that the body skeleto cossts of lkages coected to jots. However, ths aer focuses o exlorg the effect of arallelzato of the NN odes o the erformace of the bottom-u huma ose estmato. The etwork s traed to estmate the locato ad oretato of each lk ad the locato of jots. Greedy arsg s the used to assemble the lkages to huma dvduals. The algorthm has bee traed ad tested usg the OO 07 erso keyot dataset. PROPOSED NETWORK ARHITETURE The roosed system models the huma ose usg a 3- lk, 4 jot skeleto model as show Fg.. The method reseted ths aer cossts of two tasks: () body art detecto, ad () ose regresso. The body arts are deoted w h 3 w h 3 by a set P P, P,..., PN, where P ad N s a set to 3 reresetg the umber of arts detected by the roosed eural etwork. Each art P cossts D vector felds of the lk L ad a cofdece ma of two h w h w assocated jots J, where L ad J. The two assocated jots are deoted as hgh jot J h ad low jot J l searately. Jh deotes the jot that s coected wth a former body art whle J l deotes the jot that s coected wth a the later body art; more secfcally, J h h w 0, J l h w. For examle, rght shoulder - rght 0 elbow body art cossts of two assocated jots J, rght shoulder jot ad rght elbow jot ad oe lk L that coects these two jots. I ths body art, rght shoulder jot s defed as hgh jot J h whle rght elbow jot s the low jot J l. I the body art detecto task, the roosed eural etwork takes oe sgle 3-chael color mage wth sze (h w 3) as the ut ad roduces the D locato ad oretato tesor of the sze (h w 39) for each body art of ersos vsble the ut mage. The ose regresso rocess the assembles the body arts for all dvduals the ut mage usg greedy ferece. Overall Archtecture Desg: As show Fg., the roosed eural etwork archtecture ca be dvded to 3 stages: rerocessg stage, redcto stage, ad redcto stage. I the rerocessg Prerocessg Image I VGG9 F Predcto Stage Predcto Stage NN Esemble NN Esemble P P Loss NN Esemble NN Esemble P P Loss OUT NN Esemble 3 P 3 NN Esemble 3 P 3 Fgure. Proosed 3-lk huma skeleto model. Stage NN Esemble Stage NN Esemble 3x3 3x3 3x3 x x x h x w x Fgure. Archtecture of the -stage 3-arallel NN esemble etwork. h x w oyrght 08 ASME Dowloaded From: htts://roceedgs.asmedgtalcollecto.asme.org o /3/08 Terms of Use: htt://

3 stage, the color ut mage, I, s fed to a re-traed VGG9 etwork [6] to obta feature ma F. I each redctg stage, there exst 3 NN esembles to redct the lk felds ad jot cofdece mas for each of the 3 skeletal model lkages deedetly. The frst redctg stage takes the feature ma as uts ad roduces D locato ad oretato tesor, whch ca be deoted as P ( F), {,...,3}. The oututs of the frst redcto stage merge wth the feature mas F to geerate oe sgle tesor of sze h w 67 whch the serves as the ut for the secod redcto stage. The secod redcto stage s outut ca be deoted as P ( P, F), {,...,3}. The deedet archtecture of each brach s amed at achevg a deedet redcto behavor for each lkage. The mergg of the stage oe outut wth the orgal feature mas to be fed to stage two s amed at rovdg a referece to stage two for refg the lkage redctos of each body art. The oututs of each brach reset the osto ad oretato of a body art as a tesor of sze h w 3, whose last dmeso reresets the umber of chaels. The frst ad secod chaels reset the X ad Y comoet of the PAF of the lk, resectvely. The thrd chael s the heat ma of the two assocated jots. Suervso s rovded at the ed of each stage. To tra the etwork to detect all the thrtee body lkages, the loss fucto for each art has bee defed by cororatg the cofdece of the assocated jots ad the PAF for each lkage. I mult-erso mages, the et loss fucto of the etwork cossts of cotrbuto from dvdual labelled oses. Based o the olcy aled by [5], weghted fuctos are used to comute the total loss, f, as follows: f W f () where W s the bary mask dcator. It s zero whe the aotato to the th erso s mssg. Therefore, W hels avod ealzg the eural etwork whe the groud truth s mssg the dataset. The total loss fucto of the th erso, f, whch cludes the loss fucto of all the body arts s where, * J, * H J, ad * L are the loss fuctos defed for the L hgh jots, low jots, ad lks of the th body, resectvely. Body Part Presetato: To evaluate the loss fucto, a three chael-groud truth for each body art was geerated for each mage xel usg the aotated D keyots from the OO dataset. The groud truth label s geerated for each mage xel wherever ay body art elemet s vsble. The locato ad oretato of each lk s rereseted usg PAF whle the jot locato s exressed usg a bolar Gaussa jot cofdece ma. The hgh jot has a ostve cofdece ma whle the low jot uses a egatve cofdece ma to dstgush betwee each other. Let xjh,, xjl, be the groud truth of the hgh jot ad the low jot of the th body art for the th erso. The values of s jh,, s jl, for ths jot at ay xel locato x are calculated as follows: s x xjh, x x jl, jh, ex, s jl, ex where the stadard devato, σ, cotrols the sread of the eaks. The aggregato of the hgh jot s ad the low jot s groud truth are the maxmum absolute values of the dvduals, wth the assumto that o hgh jot for ay lkage cocdes wth ts corresodg lower jot. Fgure 3 shows the bolar Gaussa jot cofdece ma for the rght kee-rght akle lkage. To get a strog oretato ad osto reresetato of the body arts, PAF [5] are used to rereset the body art lkages. The D vector feld for the rght kee-rght akle body lkage s show Fg. 3. The PAF groud truth of body art of the dvdual,, * L, at ay xel x deeds o whether or ot the xel s o the art lk defed by a rego alog the vector from the hgh jot to the low jot s defed (4) f fj fl ( fjh fjl fl ) () where f J ad f L are the loss fuctos for the jot s cofdece ma ad lk PAF, resectvely. Due to the dfferet roertes of the two assocated jots, f s further dvded to J f JH ad f for the hgh ad low jots of the th body art. The JL loss fuctos for the jots ad the lks are reseted as follows: * * f,,, JH JH JH f JL JL, JL, * fl L L (3) Fgure 3. Jot cofdece ad PAF for rght keerght akle body lkage 3 oyrght 08 ASME Dowloaded From: htts://roceedgs.asmedgtalcollecto.asme.org o /3/08 Terms of Use: htt://

4 resectve ersos. The ose estmato rocess takes lace two cosecutve stes, () Body Parts Parsg: assemble art lk wth ts assocated jots, ad () Idvdual Parsg: assemblg all body arts to dvdual body skeleto. wth legth l ad wdth σ xels. The PAF vector feld * L s comuted as follows: x jl, x jh, v * L x jl, x jh, 0 0 v ( x x jh, ) l, ad 0 v ( x x jh, ) otherwse (5) Body Parts Parsg: I ths rocess, the assocato of the hgh jots, low jots, ad body art lks are determed by comutg the tegral of the PAF betwee the two jot caddates. A votg mechasm has bee used to fd the best jot-lk ars. For the hgh jot where, v s the ut D vector eredcular to v. If more tha oe erso s vsble a mage, the the PAF groud truth of ay body-art the mage s the average of all the PAFs from the vsble ersos * L * L ad low jot ar caddate ostos, x jh, m, x jl,, the assocated votg score ca be calculated as follows: (6) Vm, u u 0 where deotes the umber of body arts detected the ut mage. L x u x jh, x jl,m x jh, x jl,m du (7) x u u x jh, u x jl,m where x(u) deotes terolated osto betwee the hgh jot ad low jot. The ut to the th body art arsg s a set of the hgh jots eaks sets X jh, low jots eaks sets X jl ad System Trag: The roosed model was mlemeted usg the Keras framework [7] wth TesorFlow [8] as the backed ege. Multle stochastc gradet descet otmzers are used to otmze the eural etwork wth a total; eoch of 43. The learg rate for the etwork trag was set to 0-4. The system trag erformace s show Fg. 4 the form of decay system loss (as defed by Eq. ()) wth trag eochs. The model was traed wth 5,597 mage samles from the OO dataset o a Itel Xeo (6-ores,.8 GHz) workstato wth 3GB RAM ad NVIDIA GTX080 GPU. The trag rocess lasted 9 hours ad delvered a loss of jot cofdece ma of eochs. PAF sets L, whch ca be exressed as { X jh, X jl, L }. The outut of the body art arsg s a set of hgh jot ad low jot ars, {( x jh,, x jl, m ) : {,..., N Pjh }, m {,..., M Pjl }}. For each art, a varable m, {0,} s used to dcate whether the hgh jot ad low jot ars x jh,m, x jl, are coected or (A) (B) () (D) HUMAN POSE ESTIMATION After the cofdece ma of body arts has bee geerated by the traed etwork, the huma body ose estmato rocess takes charge of assocatg body arts to ther Fgure 5. Body art arsg: (A) Estmated hgh jot locato, (B) Estmated low jot locato, () Both Jots after body art arsg, (D) Skeleto outut after dvdual arsg Fgure 4. Model trag erformace 4 Dowloaded From: htts://roceedgs.asmedgtalcollecto.asme.org o /3/08 Terms of Use: htt:// oyrght 08 ASME

5 ot. The body arsg rocess ams to maxmze the total votg score for coectg jot ars va the followg: max VP max Vm, m, m, m, m P jl the two body arts. The ew coected jot osto ca be refed usg the followg exresso: x (8) m {,..., M }, m, jl,l x q jh, h J x jl,l x jl,l J q x qjh,h x qjh,h J x jl,l J q x qjh,h. () {,..., N Pjh }, m, m Oce commo jots are located, the coected lks ca be detfed as D huma ose skeletos. The equato esures that o jots are used to costruct more tha oe body art. EXPERIMENTS AND RESULTS Idvdual Parsg: After the body art arsg, the dvdual arsg assembles all the body arts to form dvdual skeletos by Results o OO database The OO huma keyot dataset cotas 57k mages, out of whch aroxmately 5k samles were used for the trag dataset ad the remag 5k samles were used for valdato. Durg the trag of the eural etwork, the testg that was doe o the valdato set s erformed at the ed of each eoch. Fgure 6 shows some samles from the valdato set rocessed o the fully traed etwork. A good match the redcted ad the estmated ose was observed. As reorted the trag secto, a loss of 37.7 was obtaed the jot cofdece ma geerato. assemblg rules Al,h {0,}. These rules dcate whether,q the low jot of ca be coected to the hgh jot of q ad the merged as oe jot. Aother varable store whether the q low jots of, x jl are merged wth the hgh jots of q, x jh. A votg score Vl,h,q s the used to determe the ossblty of the two jots mergg to oe jot, Vl,h,q J x jl,l J q x qjh,h x jl,l x qjh,h (9) Rutme Aalyss The rutme of the roosed algorthm comrses of two major comoets: () the body arts detecto rocess tme, whch s varat to the umber of ersos show the mage, wth rutme comlexty of O(); () the body arts Here, the egatve sg s assged to geerate ostve scores due to the egatve values of the redcted cofdece ma of the low jots. The score wll decrease as the jots grow further away. The oly exceto to ths rule ales to the eck jots where fve dfferet body lkages get coected wth ther hgh jots, Vh,h,q ' J x jh,h J q x qjh,h x jh,h x qjh,h (0) The multle-erso ose arsg rocess becomes oe otmzato roblem to maxmze the total assemblg score, max VP,Q max (Vl,h,q Al,h,q l,h,q 'Vh,h,q Ah,,hq h,,hq ),q, q l, h l, h l h Al,h,q, l {,..., M Pjl }, l,h,q, () h Al,h,q, h {,..., N Qjh }, l,h,q l Here, VP,Q resets the ossblty that jots could be merged to oe sgle jot. The coected jot osto s the refed by weghtg the redcted osto of the jots from Fgure 6. Pose Estmato results o OO 07 huma keyot valdato dataset 5 Dowloaded From: htts://roceedgs.asmedgtalcollecto.asme.org o /3/08 Terms of Use: htt:// oyrght 08 ASME

6 assemblg rocess tme, whose rutme comlexty s O( ). Hece. the rutme creases wth umber of ersos () the ut mage. omared to the body arts assemblg rocess tme, the body arts detecto rocess tme flueces the total rocessg tme eve more. Wth a lesser umber of sequetal stages comared to [5], the roosed etworks better take advatage of arallel comutg ad estmates the huma ose a more effcet way. By usg a sgle GTX080, the NN takes 03.5 ms, comared to the NN havg take 99.6 ms [5]. The tme of NN comutato decreases to 76.8 ms by alyg the etwork o two GTX080s. The body arsg takes 0.6 ms ad does ot chage wth the umber of GPUs deloyed. ONLUSION AND FUTURE WORK Ths aer reseted oe arallel, mult-brach dee eural etwork archtecture alog wth a corresodg ostrocessg method for multle erso D-ose estmato the moocular mage. Each traed brach (NN esemble) of the roosed eural etwork was traed to detect oe secfc body art (assocated jots ad lk) for the huma skeleto model ad delvered a overall detecto loss of The hghly arallel archtecture acheves smlar erformace comared to the revous dee eural etwork archtecture [5] but rus faster o a arallel comutg devces such as GPU. Ths saves tme for ost-rocessg ad the esug hgher level tasks. Ths eural etwork s also hghly adatable for tasks amed at secfc body arts sce braches of the last stage ca be deloyed deedetly. Ths feature also hels save storage ad comutg resources wthout sacrfcg erformace. The roosed system wll be augmeted wth multsectral magery to eable the detecto of huma oses varous lghtg codtos. Efforts are beg made to mrove the etwork archtecture ad ost-rocessg algorthms to acheve better effcecy ad faster rutme for uroses of deloymet o embedded systems. The roosed system wll be deloyed o a autoomous moble robotc latform to assst wth the search ad rescue of casualtes dsaster maagemet ad war-lke scearos. AKNOWLEDGMENT Ths work s suorted art by the US Army Medcal Research & Materel ommad s Telemedce & Advaced Techology Research eter (TATR), uder otract No. W8XWH The vews, oos, ad/or fdgs cotaed ths reort are those of the authors ad should ot be costrued as a offcal Deartmet of the Army osto, olcy, or decso uless so desgated by other documetato. REFERENES [] Gog, W., Zhag, X., Gozàlez, J., Sobral, A., Bouwmas, T., Tu,., ad Zahzah, E., 06, Huma Pose Estmato from Moocular Images: A omrehesve Survey, Sesors (Basel)., 6(),. 39. [] Toshev, A., DeePose: Huma Pose Estmato va Dee Neural Networks. [3] Tomso, J., Ja, A., Leu, Y., ad Bregler,., 04, Jot Trag of a ovolutoal Network ad a Grahcal Model for Huma Pose Estmato, Adv. Neural If. Process. Syst., [4] L, S., Lu, Z.-Q., ad ha, A. B., 04, Heterogeeous Mult-Task Learg for Huma Pose Estmato wth Dee ovolutoal Neural Network. [5] Agarwal, A., ad Trggs, B., 006, A Local Bass Reresetato for Estmatg Huma Pose from luttered Images, Lect. Notes omut. Sc. (cludg Subser. Lect. Notes Artf. Itell. Lect. Notes Boformatcs), 385 LNS, [6] Jag,., ad Jug, K., 008, Huma Pose Estmato Usg Actve Shae Models, Proc. World Acad. Sc., (8), [7] Shotto, J., Grshck, R. B., Ftzgbbo, A., Shar, T., ook, M., Foccho, M., Moore, R., Kohl, P., rms, A., Kma, A., ad Blake, A., 0, Effcet Huma Pose Estmato from Sgle Deth Images., IEEE Tras. Patter Aal. Mach. Itell., 35(), [8] Ye, M., Wag, X., Yag, R., Re, L., ad Pollefeys, M., 0, Accurate 3D Pose Estmato from a Sgle Deth Image, IEEE It. of. omut. Vs., [9] Iqbal, U., Mla, A., ad Gall, J., 06, PoseTrack: Jot Mult-Perso Pose Estmato ad Trackg. [0] Dalal, N., ad Trggs, W., 004, Hstograms of Oreted Gradets for Huma Detecto, 005 IEEE omut. Soc. of. omut. Vs. Patter Recogt. VPR05, (3), [] Adrluka, M., Roth, S., ad Schele, B., 009, Pctoral Structures Revsted Peole Detecto ad Artculated Pose Estmato, omuter Vso ad Patter Recogto, 009, [] Gkoxar, G., Harhara, B., Grshck, R., ad Malk, J., 04, R-NNs for Pose Estmato ad Acto Detecto,. 8. [3] Fragkadak, K., Leve, S., Felse, P., ad Malk, J., 05, Recurret Network Models for Huma Dyamcs, 05 IEEE It. of. omut. Vs., [4] Luo, P., Wag, X., ad Tag, X., 03, Pedestra Parsg va Dee Decomostoal Network, Proc. IEEE It. of. omut. Vs., [5] Datoe, M., Gall, J., Lester,., ad Va Gool, L., 03, Huma Pose Estmato Usg Body Parts Deedet Jot Regressors, 03 IEEE of. omut. Vs. Patter Recogt., [6] Ouyag, W., hu, X., ad Wag, X., 04, 4. Mult- Source Dee Learg for Huma Pose Estmato, 6 oyrght 08 ASME Dowloaded From: htts://roceedgs.asmedgtalcollecto.asme.org o /3/08 Terms of Use: htt://

7 04 IEEE of. omut. Vs. Patter Recogt., [7] he, X., ad Yulle, A., 04, Artculated Pose Estmato by a Grahcal Model wth Image Deedet Parwse Relatos,. 9. [8] We, S. E., Ramakrsha, V., Kaade, T., ad Shekh, Y., 06, ovolutoal Pose Maches, Proceedgs of the IEEE omuter Socety oferece o omuter Vso ad Patter Recogto, [9] Isafutdov, E., Pshchul, L., Adres, B., Adrluka, M., ad Schele, B., 06, Deeercut: A Deeer, Stroger, ad Faster Mult-Perso Pose Estmato Model, Lect. Notes omut. Sc. (cludg Subser. Lect. Notes Artf. Itell. Lect. Notes Boformatcs), 990 LNS, [0] Iqbal, U., ad Gall, J., 06, Mult-Perso Pose Estmato wth Local Jot-to-Perso Assocatos, Lect. Notes omut. Sc. (cludg Subser. Lect. Notes Artf. Itell. Lect. Notes Boformatcs), 994 LNS, [] Fag, H. S., Xe, S., Ta, Y. W., ad Lu,., 07, RMPE: Regoal Mult-Perso Pose Estmato, Proc. IEEE It. of. omut. Vs., 07 Octob, [] Paadreou, G., Zhu, T., Kaazawa, N., Toshev, A., Tomso, J., Bregler,., ad Murhy, K., 07, Towards Accurate Mult-Perso Pose Estmato the Wld, [3] Re, S., He, K., Grshck, R., ad Su, J., 07, Faster R-NN: Towards Real-Tme Object Detecto wth Rego Proosal Networks, IEEE Tras. Patter Aal. Mach. Itell., 39(6), [4] J, S., Ma, X., Ha, Z., Wu, Y., Yag, W., Lu, W., Qa,., ad Ouyag, W., Towards Mult-Perso Pose Trackg : Bottom-u ad To-dow Methods, (), [5] ao, Z., Smo, T., We, S.-E., ad Shekh, Y., 06, Realtme Mult-Perso D Pose Estmato Usg Part Affty Felds. [6] Smoya, K., ad Zsserma, A., 04, Very Dee ovolutoal Networks for Large-Scale Image Recogto,. 4. [7] hollet, F., ad Others, 05, Keras. [8] Abad, M., Barham, P., he, J., he, Z., Davs, A., Dea, J., Dev, M., Ghemawat, S., Irvg, G., Isard, M., Kudlur, M., Leveberg, J., Moga, R., Moore, S., Murray, D. G., Steer, B., Tucker, P., Vasudeva, V., Warde, P., Wcke, M., Yu, Y., ad Zheg, X., 06, TesorFlow: A System for Large-Scale Mache Learg, Proceedgs of the th USENIX oferece o Oeratg Systems Desg ad Imlemetato, USENIX Assocato, oyrght 08 ASME Dowloaded From: htts://roceedgs.asmedgtalcollecto.asme.org o /3/08 Terms of Use: htt://