TPT- Bench Da ase
Ti le: TPT- Bench: A L a ge- Scale, L ong- Te m andRobo - Egocen icDa ase o Benchma kingTa ge
Pe sonT acking
A u h o s: H anjingY e, Y uZ han, W eixiSi u, G uangchengChen, JingwenY u, Z iqiZ hao, K uanqiCai,
A ash A joudani, andH ongZ hang*
1 . O e iew
Thisdocumen se esas heo icialda a epo o TPT- Bench , ala ge- scaleda ase designed o Ta ge
Pe sonT acking( TPT) inc owdedanduns uc u eden i onmen s. Theda ase emphasizeslong- e m
ackingchallenges, including equen occlusions, a ge disappea ance, and e- iden i ica ion
equi emen s om a obo - egocen icpe spec i e.
Thisda ase suppo s he esea ch pape :
" TPT- Bench: A L a ge- Scale, L ong- Te m andRobo - Egocen icDa ase o Benchma kingTa ge Pe son
T acking" .
2 . E h icalS a emen and IR B A pp o al
Theda acollec ionin ol edhumansubjec sinpublicspaces.
●C onsen : Explici in o medconsen wasob ained om hep ima ysubjec s( a ge pe sons) be o e
da acollec ion.
●P i acy: N o pe sonallyiden i iablein o ma ion( PII) o bys ande swasin en ionally eco ded o
iden i ica ionpu poses. Thecollec ionadh e es o e hicalguidelines o da acollec edinpublic
spaces.
●IR B A pp o al: Thiss udy ecei ede hicalapp o al om heSou h e nU ni e si yo Scienceand
Tech nologyIR B ( A pp o alN umbe : # 2 0 2 5 JSJ1 2 8 ) onA p il0 1 , 2 0 2 5 .
3 . Da aO iginand A cquisi ion
3 . 1 C ollec ionPla o m
Theda awascollec edusingacus om- designedsenso - equippedpushingca ha mimics hekinema ics
o adi e en ial- d i emobile obo . Thesenso sui eincludes:
●Pano amicC ame a: RICO H The aZ 1 ( 3 6 0 Ddual isheye) , 1 9 2 0 x9 6 0 esolu ion@ 3 0 H z.
●Dep h C ame a: Z ED2 S e eo Came a, 6 4 0 x4 8 0 esolu ion@ 3 0 H z ( p o idingRG B- DandV isual-
Ine ialO dome y) .
●3 DL iDA R : O us e O S1 - 6 4 , 6 4 x2 0 4 8 esolu ion@ 1 0 H z.
●IM U : Buil - inIM U om heO us e L iDA R.
Senso moun ingheigh s( app oxima e) :
●Pano amicCame a: 7 2 cm
●L iDA R: 5 4 cm
●S e eo Came a: 4 5 cm
3 . 2 Scena ios
Da acollec ion ookplaceindi e seindoo andou doo en i onmen s, ca ego izedin o 1 0 high- le el
scena ios:
1 . Re ail& Shopping( e. g. , supe ma ke s, malls)
2 . F ood& Dining( e. g. , can eens, es au an s)
3 . Co ido s& H allways
4 . Roadways& W alkways
5 . V e icalT anspo ( Ele a o s/ Escala o s)
6 . O u doo Squa es& Plazas
7 . Educa ionalRooms
8 . T anspo H ubs
9 . En e ainmen & L eisu e
1 0 . L abo a o y& O ice
4 . Da ase C on en and O ganiza ion
Theda ase comp ises4 8 sequenceswi h a o aldu a iono app oxima ely5 . 3 hou s. Theda ais
o ganizedin o he ollowingdi ec o ys uc u e:
TPT- Bench/
├── pano amic_images/
│ └── < seq_id> / # e. g. , 0 0 0 1 , 0 0 0 2 . . .
│ └── < imes amp> . jpg # Equi ec angula p ojec ionimages
├── G Ts/
│ └── < seq_id> . json # G oundT u h anno a ions
├── osbags/
│ └── < seq_id> . bag # Raw senso da a( L iDA R, IM U , Images)
├── desc ip ions/
│ └── < seq_id> . x # N a u allanguagedesc ip iono hesequence
├── quick iew_ ideos/
│ └── < seq_id> . mp4 # L ow- es isualiza ion o quickinspec ion
└── e alua ion_ esul s/
└── < seq_id> /
└── < baseline> . json # O u pu om baseline acke s
4 . 1 Da aF o ma s
●Images: JPG o ma , equi ec angula p ojec ion.
●R aw Da a: RO S Bag o ma ( . bag) , con ainingsynch onizeds eams om allsenso s.
●A nno a ions: JSO N o ma .
5 . A nno a ionand M e ada aS anda ds
5 . 1 L abelingP ocess
Theda ase employsasemi- au oma icanno a ionpipeline o ensu ee iciencyandaccu acy:
1 . Ini ialEs ima ion: S a e- o - he- a " M O T + Ta ge - ReID" algo i hmswe eapplied o gene a e
ini ial acking esul s.
2 . M anualR e inemen : H umananno a o s e iewed amesa 1 5 H z. A nno a o sco ec edbounding
boxes, handledocclusions, and e i ied a ge iden i y.
3 . In e pola ion: A nno a ionswe eupsampled o 3 0 H z usinglinea in e pola ion.
4 . V e i ica ion: A inalpasswasconduc ed o e i yconsis encyac osssequences.
5 . 2 M e ada aSch ema( JSO N )
Each < seq_id> . json ilecon ains ame- le elme ada akeyedby imes amp:
{
" < imes amp> " : {
" is_exis " : 1 , / / 1 = Ta ge p esen , 0 = Ta ge absen / ullyoccluded
" is_behind_glass" : 0 , / / 1 = Behindglass( a ec sdep h) , 0 = N o behindglass, - 1 = A bsen
" bbox" : [u, , w, h], / / 2 DBoundingBox: [ op- le - x, op- le - y, wid h, heigh ]
/ / I absen : [0 , 0 , 0 , 0 ]
" in e pola ed" : 0 / / 1 = G ene a ed iain e pola ion, 0 = H umananno a ed ame
}
}
5 . 3 SequenceDesc ip ions
Each sequenceincludesa< seq_id> . x iledesc ibing:
●Ini ial a ge appea ance( clo hing, oo wea , hai ) .
●En i onmen alcondi ions( ligh ing: no mal/ dim/ da k/ b igh ) .
●E en s( clo hingchanges, speci icscena io ansi ions) .
6 . Da aIn eg i yand A u h en ici y
●V e i ica ion: A llanno a ed amesunde wen isualinspec ionby ainedanno a o s o ensu e h e
boundingboxes igh ly i he isiblepa so he a ge pe son.
●Synch oniza ion: A llmul i- modalda as eamsa e ime- synch onized ia heRobo O pe a ing
Sys em ( RO S) eco ding amewo k.
●A u h en ici y: Theda ase ep esen s eal- wo ld, unal e ed eco dingen i onmen s. N o syn he ic
agen swe eadded; allpedes iansa e ealhumansinna u alse ings.
7 . C odeand R ep oducibili y
Thecode o hebenchma ke alua ion, da a isualiza ion, andp ocessing oolki sishos edsepa a ely o
allow o e sioningandmain enance.
●C odeR eposi o y: h ps: / / gi hub. com/ M edla Tea/ TPT- BEN CH - TO O L S
●R ela ion: Thecode eposi o ycon ains hede _ oolki equi ed o pa se heJSO N anno a ionsand
p ocess he. bag iles oundin hisda ase .
8 . U sageand L icense
8 . 1 L icense
Thisda ase islicensedunde heC ea i eC ommonsA ibu ion4 . 0 In e na ional( C C BY 4 . 0 )
license. Y oua e ee o sha eandadap hema e ialaslongasyoup o ideapp op ia ec edi .
