introduction - indico-fnal (indico) · 2017-11-13 · goal for the next ~18 minutes: report the...

28
Beyond 2D representa/ons: track/shower separa/on in 3D Ji Won Park Kazu Terao 11/14/17 SLAC Na/onal Accelerator Laboratory 1

Upload: others

Post on 05-Aug-2020

0 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: INTRODUCTION - INDICO-FNAL (Indico) · 2017-11-13 · Goal for the next ~18 minutes: report the results of training a seman/c segmentaon network to perform track/shower separaon on

Beyond2Drepresenta/ons:track/showersepara/onin3D

JiWonParkKazuTerao

11/14/17

SLACNa/onalAcceleratorLaboratory1

Page 2: INTRODUCTION - INDICO-FNAL (Indico) · 2017-11-13 · Goal for the next ~18 minutes: report the results of training a seman/c segmentaon network to perform track/shower separaon on

INTRODUCTION

2

Page 3: INTRODUCTION - INDICO-FNAL (Indico) · 2017-11-13 · Goal for the next ~18 minutes: report the results of training a seman/c segmentaon network to perform track/shower separaon on

Mo/va/onsandgoals

Goalforthenext~18minutes:reporttheresultsoftrainingaseman/csegmenta/onnetworktoperformtrack/showersepara/onon3Dsimula/ondata,asaworkingtestcaseforpaSernrecogni/onin3D.

3

Whyin3D?•  Lessop/calillusionininterpre/ng3Ddata.•  PIDin2Dandtrack/showersepara/onin2Dhavebeendone

forMicroBooNEdata.PaSernrecogni/onin3Disanaturalextensionfrom2D.

Long-termmission:buildafull3Dreconstruc/onchainforLArTPCdatausingdeeplearning.

Page 4: INTRODUCTION - INDICO-FNAL (Indico) · 2017-11-13 · Goal for the next ~18 minutes: report the results of training a seman/c segmentaon network to perform track/shower separaon on

Imageclassifica3ontaskin2DFive-par/clePIDhasbeendone:givena2Dimageofasingle

par/cle,labelitasagammaray,electron,muon,pion,orproton.

4

pion muon

FromtheMicroBooNECNNpaper(2016)

gaveahappyscoredistribu3on.*Muonclassifica/onscore~highlikelythealgorithmthinksanimageisamuon.Assignedhighscoresformuonsvs.lowscoresforpionsàconfidenceinpredic/onJ

Page 5: INTRODUCTION - INDICO-FNAL (Indico) · 2017-11-13 · Goal for the next ~18 minutes: report the results of training a seman/c segmentaon network to perform track/shower separaon on

5

5-par/clePIDin3Disanaturalextension(achievedsimilarresultsas2D)

pion muon

Voxel=the3Dequivalentofpixel

Page 6: INTRODUCTION - INDICO-FNAL (Indico) · 2017-11-13 · Goal for the next ~18 minutes: report the results of training a seman/c segmentaon network to perform track/shower separaon on

METHODS

6

Page 7: INTRODUCTION - INDICO-FNAL (Indico) · 2017-11-13 · Goal for the next ~18 minutes: report the results of training a seman/c segmentaon network to perform track/shower separaon on

Ourstudy:shower/tracksepara3on.(Now3classesfortrack,shower,backgroundinsteadof5par/cleclasses)Ithasbeendonepixel-levelin2Dusingseman/csegmenta/on.

7

Truthlabel Predic/on

FromMicroBooNEDNNpaperunderreview

Yellow:track,Cyan:shower

Page 8: INTRODUCTION - INDICO-FNAL (Indico) · 2017-11-13 · Goal for the next ~18 minutes: report the results of training a seman/c segmentaon network to perform track/shower separaon on

Ourstudy:shower/tracksepara3on.(Now3classesfortrack,shower,backgroundinsteadof5par/cleclasses)Ithasbeendonepixel-levelin2Dusingseman/csegmenta/on.

8

Truthlabel Predic/on

FromMicroBooNEDNNpaperunderreview

Yellow:track,Cyan:shower

Canreusethenetworktodoshower/tracksepara/onin3D.Thisstudyallowsustoexplorehowthetechniquescalesto3D.

Page 9: INTRODUCTION - INDICO-FNAL (Indico) · 2017-11-13 · Goal for the next ~18 minutes: report the results of training a seman/c segmentaon network to perform track/shower separaon on

Theseman/csegmenta/onnetwork(SSNet)

9

Outpu

tImage

Inpu

tImage Down-sampling Up-sampling

Featuretensor

Intermediate,low-resolu/onfeaturemap

TwocomponentsofSSNet

2.Upsamplingpath1.Downsamplingpath

ThankyouKazuforthediagramJ

Page 10: INTRODUCTION - INDICO-FNAL (Indico) · 2017-11-13 · Goal for the next ~18 minutes: report the results of training a seman/c segmentaon network to perform track/shower separaon on

10

Down-sampling

Featuretensor

1.Downsamplingpath

Role:classifica3onAseriesofconvolu/onsanddownsamplingwhichreducetheinputimagedowntothelowest-resolu/onfeaturemap.Eachdownsamplingstepincreasesfieldofviewofthefeaturemapandallowsittounderstandtherela/onshipbetweenneighboringpixels.

“WriKentexts”inputimage

“Humanface”inputimage

“WriKentexts”featuremap

“Humanface”featuremap

ThankyouagainKazuforthediagramJ

Page 11: INTRODUCTION - INDICO-FNAL (Indico) · 2017-11-13 · Goal for the next ~18 minutes: report the results of training a seman/c segmentaon network to perform track/shower separaon on

11

Up-sampling

Featuretensor

2.UpsamplingpathRole:pixel-wiselabeling~reverseversionofdownsamplingpath.Aseriesofconvolu/ons-transpose,convolu/ons,andupsamplingwhichretrievetheoriginalresolu/onoftheimage,witheachpixellabeledasoneoftheclasses.

SegmentedoutputimageEachpixeliseither“human”or“background”

Page 12: INTRODUCTION - INDICO-FNAL (Indico) · 2017-11-13 · Goal for the next ~18 minutes: report the results of training a seman/c segmentaon network to perform track/shower separaon on

12

ThetypeofSSNetweused:U-ResNet1.Dow

nsam

plingpath

2.Upsam

plingpath

FeaturetensorFromMicroBooNEDNNpaperunderreview

Page 13: INTRODUCTION - INDICO-FNAL (Indico) · 2017-11-13 · Goal for the next ~18 minutes: report the results of training a seman/c segmentaon network to perform track/shower separaon on

13

ThetypeofSSNetweused:U-ResNet=U-Net+ResNet

FromMicroBooNEDNNpaperunderreview

OneResNetmodule:

WithintheU-Netarchitecture,useResNetmodules.InU-ResNet,theconvolu/onsareembeddedwithinResNetmodules.

Page 14: INTRODUCTION - INDICO-FNAL (Indico) · 2017-11-13 · Goal for the next ~18 minutes: report the results of training a seman/c segmentaon network to perform track/shower separaon on

14

ThetypeofSSNetweused:U-ResNet=U-Net+ResNet

FromMicroBooNEDNNpaperunderreview

Concatena/ons:afeatureofU-Net.Westackthefeaturemapsateachdownsamplingstagewithsame-sizefeaturemapsattheupsamplingstage.~“shortcut”opera/onstostrengthencorrela/onbetweenthelow-leveldetailsandhigh-levelcontextualinforma/on.

Page 15: INTRODUCTION - INDICO-FNAL (Indico) · 2017-11-13 · Goal for the next ~18 minutes: report the results of training a seman/c segmentaon network to perform track/shower separaon on

Genera/ngimagesforourtrainingset

15

•  3D(voxelized)•  Eachevent(image)generated

fromtruthenergydeposi/onfromLArSoi.With:–  Randomizedpar/cle

mul/plicity1~4fromauniquevertexperevent,wherethe1~4par/clesarechosenrandomlyfrom5par/cleclasses.

–  Momentumvaryingfrom100MeVto1GeVinisotropicdirec/on.

–  128x128x128voxelsà1cm^3pervoxel(forquickfirsttrial)

Proton300MeV

Proton360MeV

Electron240MeV

Pion220MeV

Inputimage:eachvoxelcontainschargeinfo.

Page 16: INTRODUCTION - INDICO-FNAL (Indico) · 2017-11-13 · Goal for the next ~18 minutes: report the results of training a seman/c segmentaon network to perform track/shower separaon on

Supervisedlearning:eachtrainingexampleisanorderedpairofinputimageandtrueoutput

image(label).

16

Yellow:track,Cyan:shower

Labelimage:eachvoxelis0(background),1(shower),or2(track).

Page 17: INTRODUCTION - INDICO-FNAL (Indico) · 2017-11-13 · Goal for the next ~18 minutes: report the results of training a seman/c segmentaon network to perform track/shower separaon on

MustweightthesoWmaxcrossentropy.Typically,animagehas99.99%background(zero-value)voxels.Evenamongnon-zerovoxels,canhaveunevennumberoftrackvoxelsvs.showervoxels.Soupweightthe“rarer”classesintheimage,e.g.ifthetruthlabelhasra/oofBG:track:shower=99:0.7:0.3,incen/vizethealgorithmtodofocusonshowermostandBGleastbyusinginversesasweights,1/99:1/0.7:1/0.3.

17

Definingtheop/miza/onobjec/ve(lossfunc/on)

Similarly,monitoralgorithm’sperformancebyevalua/ngaccuracyonlyfornon-zeropixels

Page 18: INTRODUCTION - INDICO-FNAL (Indico) · 2017-11-13 · Goal for the next ~18 minutes: report the results of training a seman/c segmentaon network to perform track/shower separaon on

Training

Op/mizer:AdamChoosebatchsizetobe8images•  batchsize~sizeofensemble,sobiggerthebeSerBUT

limitedbyGPUmemory•  oneitera/onconsumed8images

18

Page 19: INTRODUCTION - INDICO-FNAL (Indico) · 2017-11-13 · Goal for the next ~18 minutes: report the results of training a seman/c segmentaon network to perform track/shower separaon on

RESULTS

19

Page 20: INTRODUCTION - INDICO-FNAL (Indico) · 2017-11-13 · Goal for the next ~18 minutes: report the results of training a seman/c segmentaon network to perform track/shower separaon on

Non-zeropixelaccuracycurve

20

Itera/ons

Non-zeropixelaccuracy=correctlypredictednonzeropixels/totalnonzeropixelsEachitera/onconsumed8images.Lightorange:rawplotDarkorange:smoothedplot

Page 21: INTRODUCTION - INDICO-FNAL (Indico) · 2017-11-13 · Goal for the next ~18 minutes: report the results of training a seman/c segmentaon network to perform track/shower separaon on

Losscurve

21

Itera/onsEachitera/onconsumed8images.Lightorange:rawplotDarkorange:smoothedplot

Page 22: INTRODUCTION - INDICO-FNAL (Indico) · 2017-11-13 · Goal for the next ~18 minutes: report the results of training a seman/c segmentaon network to perform track/shower separaon on

Truthlabel Predic3on

22

Page 23: INTRODUCTION - INDICO-FNAL (Indico) · 2017-11-13 · Goal for the next ~18 minutes: report the results of training a seman/c segmentaon network to perform track/shower separaon on

Truthlabel Predic3on

23

Page 24: INTRODUCTION - INDICO-FNAL (Indico) · 2017-11-13 · Goal for the next ~18 minutes: report the results of training a seman/c segmentaon network to perform track/shower separaon on

Truthlabel Predic3on

24

Page 25: INTRODUCTION - INDICO-FNAL (Indico) · 2017-11-13 · Goal for the next ~18 minutes: report the results of training a seman/c segmentaon network to perform track/shower separaon on

Summaryandfuturework

WehavetrainedU-Resnettoperformshower/tracksepara/onon3Dsimula/ondataandreportatrainingaccuracyof~96%.

25

Todo:•  Exploresmallervoxelsizesforhigherprecision•  Vertexfinding(adding1moreclasstotheclassifica/ontask)•  Par/cleclustering(insteadofpixel-level,instance-awareclassifica/on)

Page 26: INTRODUCTION - INDICO-FNAL (Indico) · 2017-11-13 · Goal for the next ~18 minutes: report the results of training a seman/c segmentaon network to perform track/shower separaon on

BACKUPSLIDES

26

Page 27: INTRODUCTION - INDICO-FNAL (Indico) · 2017-11-13 · Goal for the next ~18 minutes: report the results of training a seman/c segmentaon network to perform track/shower separaon on

Overallaccuracycurve

27

Itera/ons

Page 28: INTRODUCTION - INDICO-FNAL (Indico) · 2017-11-13 · Goal for the next ~18 minutes: report the results of training a seman/c segmentaon network to perform track/shower separaon on

WhyResNet?

ThispaperdemonstrateswhyResNetissuperiortovgg,etc.inseman/csegmenta/on.

28