Changeset - 40cc3b625eb2
[Not reviewed]
default
0 2 0
Laman - 6 years ago 2019-01-22 11:18:50

work on polar hough
2 files changed with 86 insertions and 90 deletions:
0 comments (0 inline, 0 general)
exp/stone_detect.py
Show inline comments
 
import sys
 
sys.path.append("../src")
 

	
 
import os
 
import math
 
import random
 
import itertools
 

	
 
import cv2 as cv
 
import numpy as np
 
import scipy.cluster
 
import scipy.ndimage
 
import scipy.signal
 

	
 
from annotations import DataFile,computeBoundingBox
 
from hough import show
 
from analyzer.epoint import EPoint,homogenize
 
from analyzer.grid import transformPoint
 

	
 
random.seed(361)
 

	
 

	
 
class Line():
 
	def __init__(self,a,b):
 
		self.a=a
 
		self.b=b
 
		self.points={a,b}
 

	
 
	def getSortedPoints(self):
 
		return tuple(sorted(self.points))
 

	
 
	def computeAngle(self,line):
 
		ab=self.a-self.b
 
		cd=line.a-line.b
 
		alpha=math.atan(ab.y/ab.x)
 
		gamma=math.atan(cd.y/cd.x)
 
		fi=max(alpha,gamma)-min(alpha,gamma)
 
		return min(fi,math.pi-fi)
 

	
 
	def intersect(self,line):
 
		p=self.toProjective()
 
		q=line.toProjective()
 
		return EPoint.fromProjective(np.cross(p,q))
 

	
 
	def toProjective(self):
 
		return homogenize(np.cross(self.a.toProjective(),self.b.toProjective()))
 

	
 
	def transform(self,matrix):
 
		a=EPoint.fromProjective(transformPoint(self.a.toProjective(),matrix))
 
		b=EPoint.fromProjective(transformPoint(self.b.toProjective(),matrix))
 
		if a is None or b is None: return None
 
		return Line(a,b)
 

	
 
	def __str__(self): return "({0},{1})".format(self.a,self.b)
 
	def __repr__(self): return "Line({0},{1})".format(repr(self.a),repr(self.b))
 

	
 

	
 
class PolarHough:
 
	# http://www.cis.pku.edu.cn/vision/vpdetection/LiPYZ10isvc.pdf
 
	def __init__(self,anglePrecision,lengthPrecision):
 
		self._anglePrecision=anglePrecision
 
		self._lengthPrecision=lengthPrecision
 
		self._maxLength=4096
 
		n=math.ceil(2*math.pi/anglePrecision)
 
		self._acc=[[] for i in range(n)]
 

	
 
	def put(self,item):
 
		k=int(item[0]//self._anglePrecision)
 
		self._acc[k].append(item)
 

	
 
	def extract(self,count):
 
		vanishingPoints=[]
 
		angles=self._extractAngles(count)
 
		angles=[alpha for (alpha,prominence) in angles]
 
		bins=self._mapAngles(angles)
 
		for (alpha,bin) in zip(angles,bins):
 
			(length,sampleCount)=self._extractLength([dist for (beta,dist) in bin])
 
			vanishingPoints.append((alpha,length))
 
		return vanishingPoints
 

	
 
	def _extractAngles(self,k):
 
		lens=np.array(list(map(len,self._acc)))
 
		print(lens)
 
		(peakKeys,info)=scipy.signal.find_peaks(lens,prominence=0)
 
		res=sorted(zip(info["prominences"],peakKeys),reverse=True)[:k]
 
		res=[(key*self._anglePrecision,prominence) for (prominence,key) in res]
 
		print("(angle, prominence):",res,"...",[alpha/self._anglePrecision for (alpha,_) in res])
 
		return res
 

	
 
	def _mapAngles(self,angles):
 
		res=[[] for alpha in angles]
 
		for (i,bin) in enumerate(self._acc):
 
			beta=i*self._anglePrecision
 
			key=min(zip(map(lambda alpha: angleDiff(alpha,beta), angles), itertools.count()))[1]
 
			res[key].extend(bin)
 
		return res
 

	
 
	def _extractLength(self,arr):
 
		acc=np.zeros(self._maxLength+1,dtype=np.int32)
 
		for dist in arr:
 
			dist=min(dist,self._maxLength)
 
			acc[int(dist//self._lengthPrecision)]+=1
 
		res=acc.argmax()
 
		print("(length, count):",(res*self._lengthPrecision,acc[res]))
 
		return (res*self._lengthPrecision,acc[res])
 

	
 

	
 
def angleDiff(alpha,beta):
 
	diff=abs(alpha-beta)
 
	if diff>math.pi: diff=2*math.pi-diff
 
	return diff
 

	
 

	
 
def kmeans(img):
 
	arr=np.reshape(img,(-1,3)).astype(np.float)
 
	colors=np.array([[0,0,0],[255,255,255],[193,165,116]],np.float)
 
	print(colors)
 
	(centers,distortion)=scipy.cluster.vq.kmeans(arr,colors)
 
	print("k-means centers:",centers)
 
	return centers
 

	
 

	
 
def quantize(img,centers):
 
	origShape=img.shape
 
	data=np.reshape(img,(-1,3))
 
	(keys,dists)=scipy.cluster.vq.vq(data,centers)
 
	pixels=np.array([centers[k] for k in keys],dtype=np.uint8).reshape(origShape)
 
	return pixels
 

	
 

	
 
def filterStones(contours,bwImg,stoneDims):
 
	contourImg=cv.cvtColor(bwImg,cv.COLOR_GRAY2BGR)
 
	res=[]
 
	for (i,c) in enumerate(contours):
 
		keep=True
 
		moments=cv.moments(c)
 
		center=(moments["m10"]/(moments["m00"] or 1), moments["m01"]/(moments["m00"] or 1))
 
		area=cv.contourArea(c)
 
		(x,y,w,h)=cv.boundingRect(c)
 
		if w>stoneDims[0] or h>stoneDims[1]*1.5 or w<2 or h<2:
 
			cv.drawMarker(contourImg,tuple(map(int,center)),(0,0,255),cv.MARKER_TILTED_CROSS,12)
 
			keep=False
 
		coverage1=area/(w*h or 1)
 
		hull=cv.convexHull(c)
 
		coverage2=area/(cv.contourArea(hull) or 1)
 
		if coverage2<0.8:
 
			cv.drawMarker(contourImg,tuple(map(int,center)),(0,127,255),cv.MARKER_DIAMOND,12)
 
			keep=False
 
		if keep:
 
			res.append((EPoint(*center),c))
 
			cv.drawMarker(contourImg,tuple(map(int,center)),(255,0,0),cv.MARKER_CROSS)
 
	print("accepted:",len(res))
 
	print("rejected:",len(contours)-len(res))
 
	show(contourImg)
 
	return res
 

	
 

	
 
def point2lineDistance(a,b,p):
 
	# https://en.wikipedia.org/wiki/Point-line_distance#Line_defined_by_two_points
 
	ab=b-a
 
	num=abs(ab.y*p.x - ab.x*p.y + b.x*a.y - a.x*b.y)
 
@@ -145,188 +203,118 @@ def scoreMatrix(matrix,p,r,lines):
 
		if alpha<=math.pi/30:
 
			score+=alpha
 
		else: score+=math.pi/2
 
	return score
 

	
 

	
 
def groupParallels(lines,tolerance,w):
 
	ctrl=False
 
	for (i,p) in enumerate(lines):
 
		for (j,q) in enumerate(lines[i+1:]):
 
			a=p.intersect(q)
 
			if a.y>0 and a.x<w: continue
 
			for r in lines[i+1+j+1:]:
 
				b=r.intersect(p) # !! ideal points
 
				c=r.intersect(q)
 
				ab=b-a
 
				ac=c-a
 
				if abs(ab.x*ac.y-ab.y*ac.x)/2<=tolerance: # area
 
					yield (p,q,r)
 
					ctrl=True
 
					break
 
			if ctrl: break
 
		if ctrl:
 
			ctrl=False
 
			continue
 

	
 

	
 
if __name__=="__main__":
 
	filepath=sys.argv[1]
 
	annotations=DataFile(sys.argv[2])
 
	filename=os.path.basename(filepath)
 
	corners=annotations[filename][0]
 
	(x1,y1,x2,y2)=computeBoundingBox(corners)
 
	(w,h)=(x2-x1,y2-y1)
 
	img=cv.imread(filepath)
 
	(x3,x4,y3,y4)=(x1+w//4,x1+3*w//4,y1+h//4,y1+3*h//4)
 
	print("x3,x4,y3,y4:",x3,x4,y3,y4)
 
	rect=img[y3:y4,x3:x4,:]
 
	centers=kmeans(rect)
 
	print("x1,x2,y1,y2:",(x1,x2,y1,y2))
 
	img[y1:y2,x1:x2,:]=quantize(img[y1:y2,x1:x2,:],centers)
 
	print("image quantized")
 

	
 
	rect=img[y1:y2,x1:x2]
 
	unit=np.array([1,1,1],dtype=np.uint8)
 
	kernel=np.ones((3,3),np.uint8)
 
	maskB=cv.inRange(rect,centers[0]-unit,centers[0]+unit)
 
	maskB=cv.morphologyEx(maskB,cv.MORPH_OPEN,kernel,iterations=1)
 
	maskB=cv.erode(maskB,kernel,iterations=2)
 
	maskB=cv.erode(maskB,kernel,iterations=4)
 
	maskW=cv.inRange(rect,centers[1]-unit,centers[1]+unit)
 
	maskW=cv.erode(maskW,kernel,iterations=2)
 
	maskW=cv.erode(maskW,kernel,iterations=3)
 

	
 
	show(img,filename)
 
	show(maskB,filename)
 
	show(maskW,filename)
 
	stones=cv.bitwise_or(maskB,maskW)
 
	show(stones)
 

	
 
	stoneDims=(w/19,h/19)
 
	print("stone dims:",tuple(x/2 for x in stoneDims),"-",stoneDims)
 

	
 
	(contours,hierarchy)=cv.findContours(stones,cv.RETR_LIST,cv.CHAIN_APPROX_SIMPLE)
 
	stoneLocs=filterStones(contours,stones,stoneDims)
 

	
 
	linesImg=cv.cvtColor(np.zeros((h,w),np.uint8),cv.COLOR_GRAY2BGR)
 
	cv.drawContours(linesImg,[c for (point,c) in stoneLocs],-1,(255,255,255),-1)
 
	for (p,c) in stoneLocs:
 
		cv.drawMarker(linesImg,(int(p.x),int(p.y)),(255,0,0),cv.MARKER_CROSS)
 
	matsImg=np.copy(linesImg)
 

	
 
	lineDict=dict()
 
	minCount=min(max(math.sqrt(len(stoneLocs))-4,3),7)
 
	print("min count:",minCount)
 
	for line in groupLines([point for (point,contour) in stoneLocs],minCount,2):
 
		key=line.getSortedPoints()
 
		if key in lineDict: # we already have a line with the same incident points
 
			continue
 
		lineDict[line.getSortedPoints()]=line
 
		obsolete=set()
 
		for ab in lineDict.values():
 
			if ab is line: continue
 
			if line.points<ab.points: # == impossible
 
				del lineDict[key]
 
				break
 
			if ab.points<line.points:
 
				obsolete.add(ab.getSortedPoints())
 
		for key in obsolete: del lineDict[key]
 

	
 
	print("valid lines:",len(lineDict))
 
	lines=sorted(lineDict.values(), key=lambda ab: len(ab.points), reverse=True)
 
	res=[]
 
	for line in lines:
 
		v=line.b-line.a
 
		alpha=math.atan(v.y/v.x)
 
		res.append((round(math.pi/2-alpha if alpha>0 else math.pi/2+alpha,3),repr(line)))
 
		(xa,ya)=line.a
 
		(xb,yb)=line.b
 
		cv.line(linesImg,(int(xa),int(ya)),(int(xb),int(yb)),(255,255,0),1)
 
	res.sort()
 
	# for row in res: print(row)
 
	# linePack=[
 
	# 	Line(EPoint(174.457,166.127),EPoint(174.96,27.253)),
 
	# 	Line(EPoint(191.333,38.075),EPoint(191.485,71.227)),
 
	# 	Line(EPoint(210.117,167.092),EPoint(205.0,50.0)),
 
	# 	Line(EPoint(127.7,25.6),EPoint(120.172,179.405)),
 
	# 	Line(EPoint(127.809,58.481),EPoint(124.324,127.427)),
 
	# 	Line(EPoint(85.964,191.64),EPoint(97.68,14.477)),
 
	# 	Line(EPoint(56.447,124.662),EPoint(54.889,137.918))
 
	# ]
 
	# linePack=[
 
	# 	Line(EPoint(56.447,124.662),EPoint(139.695,128.104)),
 
	# 	Line(EPoint(288.267,74.433),EPoint(191.485,71.227)),
 
	# 	Line(EPoint(252.926,29.71),EPoint(174.96,27.253)),
 
	# 	Line(EPoint(274.412,120.07),EPoint(41.065,112.759)),
 
	# 	Line(EPoint(289.674,108.019),EPoint(26.17,100.538)),
 
	# 	Line(EPoint(240.702,107.818),EPoint(41.065,112.759)),
 
	# 	Line(EPoint(174.457,166.127),EPoint(88.192,164.5))
 
	# ]
 
	# for (i,p) in enumerate(linePack):
 
	# 	for q in linePack[i+1:]:
 
	# 		print(p.intersect(q))
 
	# show(linesImg)
 

	
 
	# parallels=[]
 
	# for (i,ab) in enumerate(lines):
 
	# 	for cd in lines[i+1:]:
 
	# 		if ab.computeAngle(cd)>math.pi/4:
 
	# 			continue
 
	# 		parallels.append((ab,cd))
 
	# print("parallel lines candidate pairs:",len(parallels))
 

	
 
	# parallels=list(groupParallels(lines,50,w))
 
	# print("parallel triples:",len(parallels))
 
	# for (p,q,r) in parallels:
 
	# 	print(p,q,r)
 
	# 	print(p.intersect(q))
 
	# 	print(p.intersect(r))
 
	# 	print(q.intersect(r))
 
	# 	print()
 
	# 	matsImg_=np.copy(matsImg)
 
	# 	for pi in (p,q,r):
 
	# 		cv.line(matsImg_,(int(pi.a.x),int(pi.a.y)),(int(pi.b.x),int(pi.b.y)),(0,255,255),1)
 
	# 	show(matsImg_)
 
	show(linesImg)
 

	
 
	# p=Line(corners[0],corners[1])
 
	# q=Line(corners[2],corners[3])
 
	# r=Line(corners[1],corners[2])
 
	# s=Line(corners[3],corners[0])
 
	#
 
	# matrix=computeRectiMatrix(p,q,r,s)
 
	# print(matrix)
 
	# for (point,contour) in stoneLocs:
 
	# 	point_=EPoint.fromProjective(transformPoint(point.toProjective(),matrix))
 
	# 	# print(point,"->",point_)
 
	# 	cv.line(matsImg,(int(point.x),int(point.y)),(int(point_.x),int(point_.y)),(0,255,255),1)
 
	# points1=np.float32([[p.x-x1,p.y-y1] for p in corners])
 
	# points2=np.float32([[0,0],[0,400],[400,400],[400,0]])
 
	# print(points1)
 
	# print(points2)
 
	# mat=cv.getPerspectiveTransform(points1,points2)
 
	# print(mat)
 
	# show(rect)
 
	# warped=cv.warpPerspective(rect,matrix,(400,400))
 
	# show(warped)
 

	
 
	# mats=[]
 
	# for (i,(p,q)) in enumerate(parallels):
 
	# 	for (r,s) in parallels[i+1:]:
 
	# 		if p is r or p is s or q is r or q is s:
 
	# 			continue
 
	# 		matrix=computeRectiMatrix(p,q,r,s)
 
	# 		score=scoreMatrix(matrix,p,r,lines)
 
	# 		mats.append((score,p,q,r,s,matrix))
 
	# mats.sort(key=lambda x: x[0])
 
	# for (score,p,q,r,s,matrix) in mats[:4]:
 
	# 	print(score,p,q,r,s,matrix)
 
	# 	matsImg_=np.copy(matsImg)
 
	# 	for ab in (p,q,r,s):
 
	# 		((xa,ya),(xb,yb))=(ab.a,ab.b)
 
	# 		cv.line(matsImg_,(int(xa),int(ya)),(int(xb),int(yb)),(255,255,0),1)
 
	# 	show(matsImg_)
 
	# for (score,p,q,r,s,matrix) in mats[-4:]:
 
	# 	print(score,p,q,r,s,matrix)
 
	# 	matsImg_=np.copy(matsImg)
 
	# 	for ab in (p,q,r,s):
 
	# 		((xa,ya),(xb,yb))=(ab.a,ab.b)
 
	# 		cv.line(matsImg_,(int(xa),int(ya)),(int(xb),int(yb)),(0,0,255),1)
 
	# 	show(matsImg_)
 
	imgSize=img.shape[:2]
 
	print("image size:",imgSize)
 
	imgCenter=EPoint(imgSize[1]/2,imgSize[0]/2)-EPoint(x1,y1)
 
	polarHough=PolarHough(math.pi/180,10)
 
	for (i,ab) in enumerate(lines):
 
		for cd in lines[i+1:]:
 
			point=ab.intersect(cd)
 
			if 0<=point.x<=w and 0<=point.y<=h: continue
 
			print(point,"->",point.toPolar(imgCenter))
 
			polarHough.put(point.toPolar(imgCenter))
 
	vanish=[EPoint.fromPolar(p,imgCenter) for p in polarHough.extract(2)]
 
	print(vanish)
 
	(a,b,c,d)=corners
 
	(p,q,r,s)=(Line(a,b),Line(b,c),Line(c,d),Line(d,a))
 
	v1=p.intersect(r)
 
	v2=q.intersect(s)
 
	print("true vanishing points:",v1,"~",v1.toPolar(imgCenter),"and",v2,"~",v2.toPolar(imgCenter))
src/analyzer/epoint.py
Show inline comments
 
import math
 

	
 

	
 
def homogenize(v):
 
	for k in v:
 
		if k!=0: return v/k
 
	return v
 

	
 

	
 
## Euclidean 2D plane point: (x,y).
 
class EPoint:
 
	def __init__(self,x,y):
 
		self._x=x
 
		self._y=y
 

	
 
	@property
 
	def x(self): return self._x
 
	
 
	@property
 
	def y(self): return self._y
 

	
 
	@staticmethod
 
	def fromProjective(point):
 
		if point.item(0)==0: return None
 
		return EPoint(point.item(1)/point.item(0),point.item(2)/point.item(0))
 

	
 
	@staticmethod
 
	def fromPolar(point,center):
 
		(alpha,length)=point
 
		x=math.cos(alpha)
 
		y=math.sin(alpha)
 
		return length*EPoint(x,y)+center
 

	
 
	def toProjective(self):
 
		return (1,self._x,self._y)
 

	
 
	def toPolar(self,center):
 
		v=self-center
 
		alpha=math.atan2(v.y,v.x)
 
		if alpha<0: alpha+=2*math.pi
 
		k=self.dist(center)
 
		return (alpha,k)
 

	
 
	def dist(self,a):
 
		return math.sqrt((self._x-a._x)**2+(self._y-a._y)**2)
 

	
 
	def __add__(self,a):
 
		return EPoint(self._x+a._x,self._y+a._y)
 

	
 
	def __sub__(self,a):
 
		return EPoint(self._x-a._x,self._y-a._y)
 

	
 
	def __mul__(self,k):
 
		return EPoint(self._x*k,self._y*k)
 

	
 
	def __rmul__(self,k):
 
		return self*k
 

	
 
	def __truediv__(self,k):
 
		return EPoint(self._x/k,self._y/k)
 

	
 
	def __floordiv__(self,k):
 
		return EPoint(self._x//k,self._y//k)
 

	
 
	def __neg__(self):
 
		return EPoint(-self._x,-self._y)
 

	
 
	def __getitem__(self,key):
 
		if key==0: return self._x
 
		elif key==1: return self._y
 
		raise IndexError(key)
 

	
 
	def __hash__(self):
 
		return hash((self._x,self._y))
 

	
 
	def __lt__(self,a): return self._x<a._x or (self._x==a._x and self._y<a._y)
 
	def __le__(self,a): return self._x<a._x or (self._x==a._x and self._y<=a._y)
 
	def __gt__(self,a): return self._x>a._x or (self._x==a._x and self._y>a._y)
 
	def __ge__(self,a): return self._x>a._x or (self._x==a._x and self._y>=a._y)
 
	def __eq__(self,a): return self._x==a._x and self._y==a._y
 
	def __ne__(self,a): return self._x!=a._x or self._y!=a._y
 

	
 
	def __str__(self): return "({0},{1})".format(round(self._x,3),round(self._y,3))
 
	def __repr__(self): return "EPoint({0},{1})".format(round(self._x,3),round(self._y,3))
0 comments (0 inline, 0 general)