%PDF-1.6 % 19 0 obj << /Length 6722 >> stream 0 g 0 G 0 g 0 G 0 g 0 G 0 g 0 G 0 g 0 G BT /F30 14.3462 Tf 149.587 666.202 Td [(Dynamic)-367(T)92(emp)-31(o)1(r)-1(a)1(l)-367(Filtering)-367(in)-366(Video)-367(Mo)-30(dels)]TJ/F28 9.9626 Tf 31.067 -35.009 Td [(F)83(uc)28(hen)-333(Long)]TJ/F18 6.9738 Tf 55.888 3.616 Td [(1)]TJ/F21 6.9738 Tf 4.469 0 Td [(?)]TJ/F28 9.9626 Tf 4.581 -3.616 Td [(,)-333(Zhaofan)-333(Qiu)]TJ/F18 6.9738 Tf 60.591 3.616 Td [(1)]TJ/F21 6.9738 Tf 4.469 0 Td [(?)]TJ/F28 9.9626 Tf 4.58 -3.616 Td [(,)-333(Yingw)27(ei)-333(P)28(an)]TJ/F18 6.9738 Tf 61.283 3.616 Td [(1)]TJ/F28 9.9626 Tf 4.47 -3.616 Td [(,)-333(Ting)-334(Y)84(ao)]TJ/F18 6.9738 Tf 46.48 3.616 Td [(1)]TJ/F28 9.9626 Tf 4.47 -3.616 Td [(,)]TJ -195.543 -11.955 Td [(Chong-W)83(ah)-333(Ngo)]TJ/F18 6.9738 Tf 72.212 3.616 Td [(2)]TJ/F28 9.9626 Tf 4.469 -3.616 Td [(,)-333(and)-334(T)84(ao)-333(M)-1(ei)]TJ/F18 6.9738 Tf 61.421 3.616 Td [(1)]TJ/F42 5.9776 Tf -145.554 -20.687 Td [(1)]TJ/F40 8.9664 Tf 8.758 -3.809 Td [(JD)-342(E)-1(xplore)-342(A)29(cadem)28(y)86(,)-343(Beijing,)-342(China)]TJ/F42 5.9776 Tf -26.083 -7.15 Td [(2)]TJ/F40 8.9664 Tf 8.758 -3.809 Td [(Singap)-29(ore)-342(Managemen)29(t)-343(Univ)29(ersit)28(y)86(,)-343(Singap)-28(ore)]TJ/F53 8.9664 Tf -58.587 -10.959 Td [({longfc.ustc,)-524(zhaofanqiu,)-525(panyw.ustc,)-525(tin)1(gyao.ustc}@gmail.com;)]TJ 77.652 -10.959 Td [(cwngo@smu.edu.sg;)-524(tmei@jd.com)]TJ 0 g 0 G /F55 8.9664 Tf -76.327 -43.813 Td [(Abstract.)]TJ 0 g 0 G /F40 8.9664 Tf 48.506 0 Td [(Video)-266(temp)-28(oral)-266(dynamics)-266(is)-266(con)29(v)28(en)29(tionally)-266(mo)-29(deled)-265(wi)-1(th)-265(3D)]TJ -48.506 -10.958 Td [(spatial-temp)-29(o)1(ral)-342(k)28(e)1(rn)-1(el)-341(or)-342(its)-341(factorized)-342(v)29(ersion)-342(comprised)-341(of)-342(2D)-341(spatial)]TJ 0 -10.959 Td [(k)29(ernel)-496(and)-495(1D)-495(temp)-28(oral)-495(k)28(ernel.)-495(The)-495(mo)-28(deling)-496(p)-28(o)28(w)29(er,)-495(nev)28(ertheless,)-495(is)]TJ 0 -10.959 Td [(limited)-466(b)28(y)-466(the)-466(\034xed)-466(windo)28(w)-466(size)-466(and)-466(static)-467(w)29(eigh)28(t)1(s)-467(of)-466(a)-466(k)29(ernel)-467(along)]TJ 0 -10.959 Td [(the)-387(temp)-28(oral)-387(dimension.)-387(The)-386(pre-determined)-387(k)29(ernel)-387(size)-387(sev)29(erely)-387(limits)]TJ 0 -10.959 Td [(the)-485(temp)-28(oral)-485(receptiv)29(e)-485(\034elds)-484(and)-485(the)-485(\034xed)-484(w)28(eigh)29(ts)-485(treat)-484(eac)28(h)-484(spatial)]TJ 0 -10.959 Td [(lo)-29(ca)1(tion)-287(acro)1(ss)-287(fram)1(es)-287(e)1(qu)-1(ally)86(,)-286(resulting)-286(in)-286(sub-optimal)-286(solu)-1(tion)-286(for)-286(long-)]TJ 0 -10.959 Td [(range)-469(temp)-29(oral)-469(mo)-29(deling)-469(in)-470(natural)-469(scenes.)-469(In)-470(this)-469(pap)-29(er,)-469(w)28(e)-469(presen)29(t)]TJ 0 -10.959 Td [(a)-437(new)-437(recip)-28(e)-437(of)-437(temp)-28(oral)-437(feature)-437(learning,)-437(namely)-437(Dynamic)-437(T)86(emp)-28(oral)]TJ 0 -10.959 Td [(Filter)-330(\050DTF\051,)-329(that)-330(no)29(v)28(elly)-330(p)-28(erforms)-330(spatial-a)29(w)28(a)1(re)-330(temp)-29(oral)-329(mo)-29(deling)-329(in)]TJ 0 -10.959 Td [(frequency)-354(domain)-355(with)-354(large)-355(tem)1(p)-29(oral)-354(receptiv)28(e)-354(\034eld.)-355(Sp)-28(eci\034cally)85(,)-354(DTF)]TJ 0 -10.958 Td [(dynamically)-418(learns)-417(a)-418(sp)-28(ecialized)-418(frequency)-417(\034lter)-418(for)-417(ev)28(ery)-417(spatial)-418(lo)-28(ca-)]TJ 0 -10.959 Td [(tion)-444(to)-445(mo)-28(del)-445(its)-444(long-range)-444(temp)-29(oral)-444(dynamics.)-445(Mean)29(while,)-445(the)-444(tem-)]TJ 0 -10.959 Td [(p)-29(oral)-361(fe)1(atur)-1(e)-360(of)-361(eac)28(h)-361(spatial)-361(lo)-28(cation)-361(is)-361(also)-361(transformed)-361(in)28(to)-361(frequency)]TJ 0 -10.959 Td [(feature)-281(sp)-29(ectrum)-282(via)-281(1D)-282(F)86(ast)-281(F)85(ourier)-281(T)85(ransform)-281(\050FFT\051.)-282(The)-281(sp)-29(ectrum)-281(is)]TJ 0 -10.959 Td [(mo)-28(d)-1(ulated)-362(b)28(y)-362(the)-363(learn)29(t)-363(frequency)-363(\034lter,)-362(an)-1(d)-362(then)-363(transformed)-362(b)-1(ac)29(k)-363(to)]TJ 0 -10.959 Td [(temp)-29(o)1(ral)-372(domain)-372(with)-372(in)29(v)28(erse)-372(FFT.)-371(In)-372(addition,)-372(to)-372(facilitate)-372(the)-371(learn-)]TJ 0 -10.959 Td [(ing)-468(of)-468(frequency)-468(\034lter)-468(in)-468(DTF,)-468(w)28(e)-468(p)-28(erform)-468(frame-wise)-468(aggregation)-468(to)]TJ 0 -10.959 Td [(enhance)-446(th)-1(e)-446(primary)-446(temp)-29(oral)-446(feature)-447(with)-446(its)-447(temp)-28(oral)-447(neigh)29(b)-29(ors)-446(b)28(y)]TJ 0 -10.959 Td [(in)28(te)1(r-frame)-391(correlation.)-390(It)-391(is)-390(feasible)-390(to)-391(plug)-390(DTF)-391(blo)-28(c)28(k)-390(in)28(to)-390(Con)29(vNets)]TJ 0 -10.959 Td [(and)-472(T)86(ransformer,)-471(yieldi)-1(ng)-471(DTF-Net)-472(and)-471(DTF-T)85(ransfo)1(rmer.)-472(Extensiv)28(e)]TJ 0 -10.959 Td [(exp)-29(erime)1(n)28(ts)-490(conducted)-490(on)-491(three)-490(datasets)-490(demonstrate)-490(the)-490(sup)-29(eriorit)29(y)]TJ 0 -10.958 Td [(of)-479(our)-479(prop)-29(osals.)-479(More)-479(remark)58(abl)-1(y)86(,)-479(DTF-T)86(ransformer)-479(ac)28(hiev)29(es)-479(an)-479(ac-)]TJ 0 -10.959 Td [(curacy)-514(of)-513(83.5%)-514(on)-513(Kinetics-400)-514(dataset.)-513(Source)-514(co)-28(d)-1(e)-513(is)-514(a)29(v)57(ailable)-514(at)]TJ 0 g 0 G /F53 8.9664 Tf 0 -10.959 Td [(https://github.com)1(/FuchenUSTC/DTF)]TJ 0 g 0 G /F40 8.9664 Tf 155.305 0 Td [(.)]TJ/F57 11.9552 Tf -183.651 -35.336 Td [(1)-1125(In)32(tro)-32(duction)]TJ/F28 9.9626 Tf 0 -25.789 Td [(Video)-415(is)-415(an)-415(electronic)-415(c)-1(a)1(rrier)-416(that)-415(records)-415(the)-415(ev)28(olution)-415(of)-415(mo)28(ving)-415(p)-28(ersons)-415(or)]TJ 0 -11.956 Td [(ob)-56(j)1(e)-1(cts.)-350(Mo)-27(de)-1(ling)-350(suc)28(h)-350(ev)28(olution)-350(with)-350(time)-350(is)-351(essen)28(tial)-350(to)-350(the)-350(understanding)-350(of)]TJ 0 -11.955 Td [(motion)-267(patterns)-268(in)-268(v)1(i)-1(deos.)-267(The)-268(recen)28(t)-267(adv)55(ances)-268(g)1(e)-1(nerally)-267(hinge)-268(on)-267(temp)-28(oral)-267(con-)]TJ 0 -11.955 Td [(v)28(olution)-224(for)-224(temp)-28(oral)-224(mo)-28(deling)-224(in)-224(video)-224(mo)-27(dels)-1(.)-224(F)84(urthermore,)-224(the)-224(common)-224(recip)-28(e)]TJ 0 -11.955 Td [(is)-402(to)-402(in)28(tegrate)-402(temp)-28(oral)-402(con)28(v)28(olution)-402(in)27(to)-402(space-time)-402(3D)-402(con)28(v)28(olution)-402([)]TJ 0 g 0 G [(16)]TJ 0 g 0 G [(,)]TJ 0 g 0 G [(50)]TJ 0 g 0 G [(])-402(or)]TJ 0 g 0 G ET q 1 0 0 1 134.765 139.538 cm []0 d 0 J 0.398 w 0 0 m 56.693 0 l S Q BT /F44 5.9776 Tf 137.339 132.887 Td [(?)]TJ/F40 8.9664 Tf 7.388 -3.809 Td [(F)86(uc)28(hen)-342(Long)-343(and)-342(Z)-1(hao)1(fan)-343(Qiu)-343(co)1(n)28(tributed)-343(equally)-342(to)-343(this)-342(w)28(ork.)]TJ 0 g 0 G 0 g 0 G 0 g 0 G ET endstream endobj 43 0 obj << /Length 6873 >> stream 0 g 0 G BT /F40 8.9664 Tf 134.765 692.105 Td [(2)-3172(F.)-343(Long,)-342(Z.)-343(Qiu,)-343(Y.)-342(P)28(a)1(n)-1(,)-342(T.)-343(Y)86(ao,)-342(C.W.)-343(Ngo)-342(and)-343(T.)-342(M)-1(e)1(i)]TJ 0 g 0 G 0 g 0 G ET 1 0 0 1 176.265 590.117 cm q .19446 0 0 .19446 0 0 cm q 1 0 0 1 0 0 cm /Im1 Do Q Q 0 g 0 G 1 0 0 1 -176.265 -590.117 cm BT /F28 9.9626 Tf 134.765 571.786 Td [(Fig.)-167(1:)-365(Mo)-28(deling)-365(the)-366(temp)-28(oral)-365(ev)28(olution)-366(of)-365(t)28(w)27(o)-365(spatial)-366(regions)-365(\050mark)28(ed)-366(in)-365(pink)]TJ 0 -11.955 Td [(and)-333(orange)-333(on)-334(the)-333(\034lmstrip\051)-333(via)-334(\050a\051)-333(temp)-28(oral)-333(con)28(v)28(olution)-334(and)-333(\050b\051)-333(our)-333(DTF.)]TJ 0 g 0 G 0 g 0 G 0 -36.231 Td [(explicitly)-324(utilize)-324(temp)-28(oral)-324(con)27(v)28(olution)-324(to)-324(co-w)28(ork)-324(with)-325(spatial)-324(con)28(v)28(olution.)-324(Fig-)]TJ 0 -11.955 Td [(ure)]TJ 0 g 0 G [-262(1)]TJ 0 g 0 G [(\050a\051)-261(conc)-1(eptually)-261(depicts)-262(the)-262(temp)-28(oral)-262(mo)-27(deling)-262(pro)-28(cesses)-262(with)-262(1D)-261(temp)-28(oral)]TJ 0 -11.955 Td [(con)28(v)28(olution)-342(for)-342(t)28(w)28(o)-342(di\033eren)28(t)-342(spatial)-342(lo)-28(cations)-341(within)-342(the)-342(input)-342(video.)-341(The)-342(tem-)]TJ 0 -11.955 Td [(p)-28(oral)-333(con)28(v)28(olution)-333(lo)-28(cally)-333(agg)1(regates)-333(the)-333(features)-333(of)-333(the)-333(same)-333(spatial)-333(lo)-28(cation)-333(in)]TJ 0 -11.955 Td [(adjacen)28(t)-348(frames.)-347(The)-348(temp)-28(oral)-347(receptiv)27(e)-347(\034eld)-348(is)-347(th)27(us)-347(\034xed)-348(and)-347(the)-348(corresp)-28(ond-)]TJ 0 -11.956 Td [(ing)-406(k)28(e)-1(rnel)-406(w)28(eigh)28(ts)-407(are)-406(the)-407(same)-406(across)-406(di\033eren)27(t)-406(spatial)-406(lo)-28(cations.)-407(This)-406(setting)]TJ 0 -11.955 Td [(inevitably)-272(limits)-272(the)-271(temp)-28(oral)-272(receptiv)28(e)-272(\034eld)-272(and)-272(ignor)1(e)-1(s)-271(the)-272(inheren)28(t)-272(di\033erences)]TJ 0 -11.955 Td [(of)-324(spatial)-324(con)28(texts)-324(at)-323(v)55(aried)-324(lo)-27(cations)-324(during)-324(temp)-28(oral)-324(mo)-27(delin)-1(g)1(.)-324(Figure)]TJ 0 g 0 G [-324(1)]TJ 0 g 0 G [-324(illus-)]TJ 0 -11.955 Td [(trates)-438(t)28(w)27(o)-438(spatial)-438(lo)-28(cations)-438(exhibiting)-438(di\033eren)28(t)-438(spatial)-438(con)27(ten)28(ts:)-438(the)-438(pink)-438(dot)]TJ 0 -11.955 Td [(refers)-284(to)-285(the)-284(trac)28(k)-285(in)-284(the)-284(static)-285(bac)28(kground,)-284(whi)-1(le)-284(the)-284(orange)-285(dot)-284(sho)28(ws)-285(a)-284(p)-28(erson)]TJ 0 -11.956 Td [(mo)28(ving)-391(rapidly)-390(across)-391(a)-391(constan)28(tly)-390(c)27(hanging)-390(bac)27(kgro)1(u)-1(nd.)-390(As)-391(the)-390(e)-1(v)28(olutions)-390(of)]TJ 0 -11.955 Td [(spatial)-319(lo)-28(cations)-319(corresp)-27(ond)-319(to)-319(motion)-319(patterns)-319(sp)-28(eci\034c)-319(to)-319(di\033eren)28(t)-319(mo)28(v)27(emen)28(ts,)]TJ 0 -11.955 Td [(an)-329(optimal)-329(w)28(a)27(y)-329(of)-329(mo)-28(deling)-329(is)-329(b)28(y)-329(ha)28(ving)-329(di\033eren)27(t)-329(k)28(ernels)-329(with)-329(v)55(arying)-329(size)-329(and)]TJ 0 -11.955 Td [(w)28(eigh)27(ts)-341(that)-342(c)28(haracterize)-342(their)-342(resp)-28(ectiv)28(e)-342(spatial)-342(con)28(text)-342(with)-341(su\036ci)-1(en)28(t)-342(temp)-27(o-)]TJ 0 -11.955 Td [(ral)-332(receptiv)28(e)-332(\034)-1(eld.)-332(Ha)28(ving)-332(the)-332(same)-332(k)27(ernel)-332(\034lters)-332(o)28(v)28(er)-332(d)-1(i\033eren)28(t)-332(spatial)-332(lo)-28(cations)]TJ 0 -11.955 Td [(will)-333(h)27(urt)-333(the)-333(mining)-333(of)-334(long-range)-333(temp)-28(oral)-333(dep)-28(endency)83(.)]TJ 14.944 -11.958 Td [(In)-235(this)-234(pap)-28(er,)-234(w)27(e)-234(prop)-28(ose)-235(to)-234(mitigate)-235(these)-234(issue)-1(s)-234(b)28(y)-235(form)28(ulating)-235(the)-234(temp)-28(oral)]TJ -14.944 -11.955 Td [(feature)-256(learning)-255(in)-256(the)-256(frequency)-255(domain,)-256(pursuing)-256(a)-255(dynamic)-256(spatial-a)28(w)28(are)-256(tem-)]TJ 0 -11.955 Td [(p)-28(oral)-295(mo)-28(deling)-295(with)-296(an)-295(enlarged)-295(temp)-28(oral)-296(receptiv)28(e)-295(\034eld.)-296(Sp)-28(eci\034cally)84(,)-296(w)28(e)-295(design)]TJ 0 -11.955 Td [(a)-252(dy)1(n)-1(a)1(m)-1(ic)-251(temp)-28(oral)-252(\034lter)-251(\050see)-252(Figure)]TJ 0 g 0 G [-251(1)]TJ 0 g 0 G [(\050b\051\051)-252(to)-252(c)28(haracterize)-252(temp)-27(oral)-252(ev)28(olution)-252(b)28(y)]TJ 0 -11.955 Td [(learning)-373(frequency)-373(\034lter)-373(to)-372(adaptiv)27(ely)-372(mo)-28(dulate)-373(the)-373(sp)-28(ectrum)-373(of)-372(tem)-1(p)-27(oral)-373(fea-)]TJ 0 -11.955 Td [(tures)-391(at)-392(di\033eren)28(t)-392(spatial)-391(lo)-28(cations.)-391(A)27(ccording)-391(to)-392(the)-391(con)28(v)28(olution)-392(theorem)-391([)]TJ 0 g 0 G [(39)]TJ 0 g 0 G [(],)]TJ 0 -11.956 Td [(the)-405(p)-28(oin)28(t-wise)-405(m)28(ultiplication)-405(of)-405(sp)-27(ectrums)-405(in)-405(frequency)-405(domain)-405(of)-404(t)27(w)28(o)-405(signals)]TJ 0 -11.955 Td [(is)-414(equiv)56(alen)28(t)-414(to)-414(the)-413(temp)-28(oral)-414(con)28(v)28(olution)-414(b)-28(et)28(w)28(een)-414(them.)-414(As)-413(s)-1(uc)28(h,)-413(c)-1(o)1(n)-1(sidering)]TJ 0 -11.955 Td [(that)-317(the)-317(learn)28(t)-317(sp)-28(ecialized)-317(frequency)-317(\034lter)-317(op)-28(erates)-317(o)28(v)28(er)-317(all)-317(the)-317(frequencies,)-317(this)]TJ 0 -11.955 Td [(frequency)-268(\034lter)-269(can)-268(b)-28(e)-268(in)28(te)-1(r)1(p)-1(r)1(e)-1(ted)-268(as)-268(a)-269(temp)-27(oral)-269(con)28(v)28(olution)-269(with)-268(a)-268(larger)-269(k)28(ernel)]TJ 0 -11.955 Td [(size.)-356(The)-356(design)-356(nicely)-356(enhances)-356(the)-356(mining)-356(of)-356(long-term)-356(temp)-28(o)1(ral)-356(dep)-28(endency)]TJ 0 -11.955 Td [(in)-374(the)-375(frequency)-374(domain,)-374(without)-375(increasing)-374(computational/memory)-375(o)28(v)28(erhead.)]TJ 0 -11.956 Td [(Mean)28(while,)-257(in)-257(an)-256(e\033ort)-257(to)-257(deal)-256(w)-1(ith)-256(di\033eren)27(t)-256(con)28(texts)-257(of)-257(spatial)-256(l)-1(o)-27(cations,)-257(w)28(e)-257(dy-)]TJ 0 -11.955 Td [(namically)-242(learn)-243(a)-242(sp)-28(ecialized)-243(frequency)-242(\034lter)-243(for)-242(eac)28(h)-243(spatial)-242(lo)-28(cation)-242(based)-243(on)-242(its)]TJ 0 -11.955 Td [(temp)-28(oral)-335(features)-336(across)-335(time)-1(.)-335(F)83(urthermore,)-335(frame-wise)-336(aggregation)-335(is)-336(uniquely)]TJ 0 -11.955 Td [(exploited)-354(to)-354(strengthen)-354(the)-355(primar)1(y)-355(temp)-27(oral)-354(feature)-355(o)1(f)-355(eac)28(h)-354(lo)-28(cation)-354(b)28(y)-354(accu-)]TJ 0 -11.955 Td [(m)28(ulating)-337(its)-337(temp)-28(oral)-337(neigh)28(b)-28(ors)-337(with)-337(in)28(ter-frame)-337(atten)28(tion,)-337(thereb)28(y)-337(facilitating)]TJ 0 -11.956 Td [(the)-333(learning)-334(of)-333(frequency)-333(\034lter.)]TJ 0 g 0 G 0 g 0 G ET endstream endobj 37 0 obj << /Type /XObject /Subtype /Form /FormType 1 /PTEX.FileName (./intro_fig.pdf) /PTEX.PageNumber 1 /PTEX.InfoDict 47 0 R /BBox [0 0 1351.5 442.5] /Group 46 0 R /Resources << /ExtGState << /GS5 48 0 R /GS6 49 0 R /GS8 50 0 R >>/XObject << /Image7 51 0 R /Image9 52 0 R /Image10 53 0 R /Image11 54 0 R /Image23 55 0 R /Image24 56 0 R /Image25 57 0 R /Image26 58 0 R >>/Font << /F1 59 0 R/F2 60 0 R/F3 61 0 R/F4 62 0 R>> /ProcSet [ /PDF /Text /ImageB /ImageC /ImageI ] >> /Length 9882 /Filter /FlateDecode >> stream x]]dq}_@gl뒼$"8 A+ǬlmV1Sdϝi'J aEC"Y=|Oo~/~v8|ˇ|?>{~&Zchv\!|lO$q^J>7fׇ