Skip to content

Commit

Permalink
initial version
Browse files Browse the repository at this point in the history
  • Loading branch information
Marco Pedersoli committed Jul 28, 2014
0 parents commit b44248b
Show file tree
Hide file tree
Showing 48 changed files with 371,454 additions and 0 deletions.
266 changes: 266 additions & 0 deletions VOCpr.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,266 @@
#from hog import *
import numpy
import pylab
from database import *
from util import box,overlap,overlapx
import time

def cmpscore(a,b):
return -cmp(a[1],b[1])


def VOCprRecord(gtImages,detlist,show=False,ovr=0.5,pixels=None):
"""
calculate the precision recall curve
"""
dimg={}
tot=0
for idx in range(len(gtImages)):
rect=gtImages[idx]["bbox"][:]
if rect!=[]:
dimg[gtImages[idx]["name"].split("/")[-1].split(".")[0]]={"bbox":rect,"det":[False]*len(rect)}
for i, recti in enumerate(rect):
if recti[5] == 0:
tot=tot+1

imname=[]
cnt=0
tp=numpy.zeros(len(detlist))
fp=numpy.zeros(len(detlist))
thr=numpy.zeros(len(detlist))
detlist.sort(cmpscore)
for idx,detbb in enumerate(detlist):
found=False
maxovr=0
gt=0
if dimg.has_key(detbb[0]):
rect=dimg[detbb[0]]["bbox"]
found=False
for ir,r in enumerate(rect):
rb=(float(detbb[3]),float(detbb[2]),float(detbb[5]),float(detbb[4]))
if pixels==None:
covr=overlap(rb,r)
else:
covr=overlapx(rb,r,pixels)
if covr>=maxovr:
maxovr=covr
gt=ir

if maxovr>ovr:
if dimg[detbb[0]]["bbox"][gt][5] == 0:
if not(dimg[detbb[0]]["det"][gt]):
tp[idx]=1
dimg[detbb[0]]["det"][gt]=True
else:
fp[idx]=1
else:
fp[idx]=1
thr[idx]=detbb[1]
if show:
prec=numpy.sum(tp)/float(numpy.sum(tp)+numpy.sum(fp))
rec=numpy.sum(tp)/tot
print "Scr:",detbb[1],"Prec:%.3f"%prec,"Rec:%.3f"%rec
ss=raw_input()
if ss=="s" or not(found):
pylab.ioff()
img=gtImages.getImageByName2(detbb[0])
pylab.figure(1)
pylab.clf()
pylab.imshow(img)
rb=(float(detbb[3]),float(detbb[2]),float(detbb[5]),float(detbb[4]))
for r in rect:
pylab.figure(1)
pylab.ioff()
box(r[0],r[1],r[2],r[3],'b',lw=1.5)
if found:
box(rb[0],rb[1],rb[2],rb[3],'g',lw=1.5)
else:
box(rb[0],rb[1],rb[2],rb[3],'r',lw=1.5)
pylab.draw()
pylab.show()
rect=[]

return tp,fp,thr,tot

def VOCprRecordOptim(gtImages,detlist,show=False,ovr=0.5,pixels=None):
"""
calculate the precision recall curve
"""
tx=[];ty=[];sx=[];sy=[]
dimg={}
tot=0
for idx in range(len(gtImages)):
rect=gtImages[idx]["bbox"][:]
if rect!=[]:
dimg[gtImages[idx]["name"].split("/")[-1].split(".")[0]]={"bbox":rect,"det":[False]*len(rect)}
for i, recti in enumerate(rect):
if recti[5] == 0:
tot=tot+1

imname=[]
cnt=0
tp=numpy.zeros(len(detlist))
fp=numpy.zeros(len(detlist))
thr=numpy.zeros(len(detlist))
detlist.sort(cmpscore)
for idx,detbb in enumerate(detlist):
found=False
maxovr=0
gt=0
if dimg.has_key(detbb[0]):
rect=dimg[detbb[0]]["bbox"]
found=False
for ir,r in enumerate(rect):
rb=(float(detbb[3]),float(detbb[2]),float(detbb[5]),float(detbb[4]))
if pixels==None:
covr=overlap(rb,r)
else:
covr=overlapx(rb,r,pixels)
if covr>=maxovr:
maxovr=covr
gt=ir

if maxovr>ovr:
if dimg[detbb[0]]["bbox"][gt][5] == 0:
if not(dimg[detbb[0]]["det"][gt]):
tp[idx]=1
dimg[detbb[0]]["det"][gt]=True
gtx=dimg[detbb[0]]["bbox"][gt][3]-dimg[detbb[0]]["bbox"][gt][1]
dtx=detbb[4]-detbb[2]
gty=dimg[detbb[0]]["bbox"][gt][2]-dimg[detbb[0]]["bbox"][gt][0]
dty=detbb[5]-detbb[3]
gtcx=(dimg[detbb[0]]["bbox"][gt][3]+dimg[detbb[0]]["bbox"][gt][1])/2.
dtcx=(detbb[4]+detbb[2])/2.
gtcy=(dimg[detbb[0]]["bbox"][gt][2]+dimg[detbb[0]]["bbox"][gt][0])/2.
dtcy=(detbb[5]+detbb[3])/2.
tx.append((gtcx-dtcx)/float(dtx))
ty.append((gtcy-dtcy)/float(dty))
sx.append(gtx/float(dtx))
sy.append(gty/float(dty))
else:
fp[idx]=1
else:
fp[idx]=1

thr[idx]=detbb[1]
if show:
prec=numpy.sum(tp)/float(numpy.sum(tp)+numpy.sum(fp))
rec=numpy.sum(tp)/tot
print "Scr:",detbb[1],"Prec:%.3f"%prec,"Rec:%.3f"%rec
ss=raw_input()
if ss=="s" or not(found):
pylab.ioff()
img=gtImages.getImageByName2(detbb[0])
pylab.figure(1)
pylab.clf()
pylab.imshow(img)
rb=(float(detbb[3]),float(detbb[2]),float(detbb[5]),float(detbb[4]))
for r in rect:
pylab.figure(1)
pylab.ioff()
box(r[0],r[1],r[2],r[3],'b',lw=1.5)
if found:
box(rb[0],rb[1],rb[2],rb[3],'g',lw=1.5)
else:
box(rb[0],rb[1],rb[2],rb[3],'r',lw=1.5)
pylab.draw()
pylab.show()
rect=[]

return tp,fp,thr,tot,tx,ty,sx,sy

def VOCap(rec,prec):
mrec=numpy.concatenate(([0],rec,[1]))
mpre=numpy.concatenate(([0],prec,[0]))
for i in range(len(mpre)-2,0,-1):
mpre[i]=max(mpre[i],mpre[i+1]);
i=numpy.where(mrec[1:]!=mrec[0:-1])[0]+1;
ap=numpy.sum((mrec[i]-mrec[i-1])*mpre[i]);
return ap

def VOColdap(rec,prec):
rec=numpy.array(rec)
prec=numpy.array(prec)
ap=0.0
for t in numpy.linspace(0,1,11):
pr=prec[rec>=t]
if pr.size==0:
pr=0
p=numpy.max(pr);
ap=ap+p/11.0;
return ap

def drawPrfast(tp,fp,tot,show=True,col="g"):
tp=numpy.cumsum(tp)
fp=numpy.cumsum(fp)
rec=tp/tot
prec=tp/(fp+tp)
ap=VOColdap(rec,prec)
ap1=VOCap(rec,prec)
if show:
pylab.plot(rec,prec,'-%s'%col)
pylab.title("AP=%.1f 11pt(%.1f)"%(ap1*100,ap*100))
pylab.xlabel("Recall")
pylab.ylabel("Precision")
pylab.grid()
pylab.gca().set_xlim((0,1))
pylab.gca().set_ylim((0,1))
pylab.show()
pylab.draw()
return rec,prec,ap1

def filterdet(det,minpix,scale=1.0):
ndet=[]
minwidth = 100000000000
minheight = minwidth
minsize = minwidth
for idl,l in enumerate(det):
minwidth = min(minwidth, l[4]-l[2])
minheight= min(minheight, l[5]-l[3])
minsize = min(minwidth, minheight)
if ((l[4]-l[2])>minpix) or ((l[5]-l[3])>minpix):
ndet.append([l[0],l[1],l[2]*scale,l[3]*scale,l[4]*scale,l[5]*scale])
print("Minimum detection size in the dataset is: ", minsize)
return ndet

def transf_dets(dets,tx,ty,sx,sy):
dets2=[]
for el in dets:
w=(el[4]-el[2])/2.0
h=(el[5]-el[3])/2.0
cx=(el[4]+el[2])/2.+tx*w*2
cy=(el[5]+el[3])/2.+ty*h*2
dets2.append([el[0],el[1],cx-w*sx,cy-h*sy,cx+w*sx,cy+h*sy])
return dets2

counter =-1
def evaluate_optim(tsImages,dets, lab,color,iter=3,point=False,ovr=0.5):
global counter
counter +=1

ttx=0;tty=0;tsx=1;tsy=1
for l in range(iter):
tp,fp,scr,tot,tx,ty,sx,sy=VOCprRecordOptim(tsImages,dets,show=False,ovr=ovr)
#for each TP compute the mean translation (tx,ty) and mean scale (sx,sy)
tx=numpy.mean(tx)
ty=numpy.mean(ty)
sx=numpy.mean(sx)
sy=numpy.mean(sy)
ttx+=tx;tty+=ty
tsx*=sx;tsy*=sy
#print "#%d tx:%f ty:%f sx:%f sy:%f"%(sum(tp),tx,ty,sx,sy)
dets=transf_dets(dets,tx,ty,sx,sy)
print "Tranformations "+lab,ttx,tty,tsx,tsy

rc,pr,ap1=drawPrfast(tp,fp,tot,show=False)
if point:
lb3 ="%s "%(lab)
plotid, = pylab.plot(rc[-1],pr[-1],color=color,label=lb3,marker="o", markersize=10 , zorder=50)
ap1 = -1
else:
lb3 ="%s (AP %.02f)" %(lab, ap1*100)
plotid, = pylab.plot(rc,pr,color=color,label=lb3,linewidth=5 , zorder=10)
return (ap1, plotid, lb3, counter)



Binary file added annotations/Annotations_Face_PASCALLayout_large.mat
Binary file not shown.
Binary file added annotations/Annotations_Face_PASCALLayout_large_fixed.mat
Binary file not shown.
Binary file added annotations/anno2.mat
Binary file not shown.
Binary file added annotations/new_annotations_AFW.mat
Binary file not shown.
Loading

0 comments on commit b44248b

Please sign in to comment.