gazesim/code/pupil/methods.py
2016-03-09 19:52:35 +01:00

661 lines
21 KiB
Python

'''
(*)~----------------------------------------------------------------------------------
Pupil - eye tracking platform
Copyright (C) 2012-2015 Pupil Labs
Distributed under the terms of the CC BY-NC-SA License.
License details are in the file license.txt, distributed as part of this software.
----------------------------------------------------------------------------------~(*)
'''
import numpy as np
try:
import numexpr as ne
except:
ne = None
import cv2
import logging
logger = logging.getLogger(__name__)
class Roi(object):
"""this is a simple 2D Region of Interest class
it is applied on numpy arrays for convenient slicing
like this:
roi_array_slice = full_array[r.view]
# do something with roi_array_slice
this creates a view, no data copying done
"""
def __init__(self, array_shape):
self.array_shape = array_shape
self.lX = 0
self.lY = 0
self.uX = array_shape[1]
self.uY = array_shape[0]
self.nX = 0
self.nY = 0
@property
def view(self):
return slice(self.lY,self.uY,),slice(self.lX,self.uX)
@view.setter
def view(self, value):
raise Exception('The view field is read-only. Use the set methods instead')
def add_vector(self,(x,y)):
"""
adds the roi offset to a len2 vector
"""
return (self.lX+x,self.lY+y)
def sub_vector(self,(x,y)):
"""
subs the roi offset to a len2 vector
"""
return (x-self.lX,y-self.lY)
def set(self,vals):
if vals is not None and len(vals) is 5:
if vals[-1] == self.array_shape:
self.lX,self.lY,self.uX,self.uY,_ = vals
else:
logger.info('Image size has changed: Region of Interest has been reset')
elif vals is not None and len(vals) is 4:
self.lX,self.lY,self.uX,self.uY= vals
def get(self):
return self.lX,self.lY,self.uX,self.uY,self.array_shape
def bin_thresholding(image, image_lower=0, image_upper=256):
binary_img = cv2.inRange(image, np.asarray(image_lower),
np.asarray(image_upper))
return binary_img
def make_eye_kernel(inner_size,outer_size):
offset = (outer_size - inner_size)/2
inner_count = inner_size**2
outer_count = outer_size**2-inner_count
val_inner = -1.0 / inner_count
val_outer = -val_inner*inner_count/outer_count
inner = np.ones((inner_size,inner_size),np.float32)*val_inner
kernel = np.ones((outer_size,outer_size),np.float32)*val_outer
kernel[offset:offset+inner_size,offset:offset+inner_size]= inner
return kernel
def dif_gaus(image, lower, upper):
lower, upper = int(lower-1), int(upper-1)
lower = cv2.GaussianBlur(image,ksize=(lower,lower),sigmaX=0)
upper = cv2.GaussianBlur(image,ksize=(upper,upper),sigmaX=0)
# upper +=50
# lower +=50
dif = lower-upper
# dif *= .1
# dif = cv2.medianBlur(dif,3)
# dif = 255-dif
dif = cv2.inRange(dif, np.asarray(200),np.asarray(256))
kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5,5))
dif = cv2.dilate(dif, kernel, iterations=2)
dif = cv2.erode(dif, kernel, iterations=1)
# dif = cv2.max(image,dif)
# dif = cv2.dilate(dif, kernel, iterations=1)
return dif
def equalize(image, image_lower=0.0, image_upper=255.0):
image_lower = int(image_lower*2)/2
image_lower +=1
image_lower = max(3,image_lower)
mean = cv2.medianBlur(image,255)
image = image - (mean-100)
# kernel = cv2.getStructuringElement(cv2.MORPH_CROSS, (3,3))
# cv2.dilate(image, kernel, image, iterations=1)
return image
def erase_specular(image,lower_threshold=0.0, upper_threshold=150.0):
"""erase_specular: removes specular reflections
within given threshold using a binary mask (hi_mask)
"""
thresh = cv2.inRange(image,
np.asarray(float(lower_threshold)),
np.asarray(256.0))
kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (7,7))
hi_mask = cv2.dilate(thresh, kernel, iterations=2)
specular = cv2.inpaint(image, hi_mask, 2, flags=cv2.INPAINT_TELEA)
# return cv2.max(hi_mask,image)
return specular
def find_hough_circles(img):
circles = cv2.HoughCircles(pupil_img,cv2.cv.CV_HOUGH_GRADIENT,1,20,
param1=50,param2=30,minRadius=0,maxRadius=80)
if circles is not None:
circles = np.uint16(np.around(circles))
for i in circles[0,:]:
# draw the outer circle
cv2.circle(img,(i[0],i[1]),i[2],(0,255,0),2)
# draw the center of the circle
cv2.circle(img,(i[0],i[1]),2,(0,0,255),3)
def chessboard(image, pattern_size=(9,5)):
status, corners = cv2.findChessboardCorners(image, pattern_size, flags=4)
if status:
mean = corners.sum(0)/corners.shape[0]
# mean is [[x,y]]
return mean[0], corners
else:
return None
def curvature(c):
try:
from vector import Vector
except:
return
c = c[:,0]
curvature = []
for i in xrange(len(c)-2):
#find the angle at i+1
frm = Vector(c[i])
at = Vector(c[i+1])
to = Vector(c[i+2])
a = frm -at
b = to -at
angle = a.angle(b)
curvature.append(angle)
return curvature
def GetAnglesPolyline(polyline,closed=False):
"""
see: http://stackoverflow.com/questions/3486172/angle-between-3-points
ported to numpy
returns n-2 signed angles
"""
points = polyline[:,0]
if closed:
a = np.roll(points,1,axis=0)
b = points
c = np.roll(points,-1,axis=0)
else:
a = points[0:-2] # all "a" points
b = points[1:-1] # b
c = points[2:] # c points
# ab = b.x - a.x, b.y - a.y
ab = b-a
# cb = b.x - c.x, b.y - c.y
cb = b-c
# float dot = (ab.x * cb.x + ab.y * cb.y); # dot product
# print 'ab:',ab
# print 'cb:',cb
# float dot = (ab.x * cb.x + ab.y * cb.y) dot product
# dot = np.dot(ab,cb.T) # this is a full matrix mulitplication we only need the diagonal \
# dot = dot.diagonal() # because all we look for are the dotproducts of corresponding vectors (ab[n] and cb[n])
dot = np.sum(ab * cb, axis=1) # or just do the dot product of the correspoing vectors in the first place!
# float cross = (ab.x * cb.y - ab.y * cb.x) cross product
cros = np.cross(ab,cb)
# float alpha = atan2(cross, dot);
alpha = np.arctan2(cros,dot)
return alpha*(180./np.pi) #degrees
# return alpha #radians
# if ne:
# def GetAnglesPolyline(polyline):
# """
# see: http://stackoverflow.com/questions/3486172/angle-between-3-points
# ported to numpy
# returns n-2 signed angles
# same as above but implemented using numexpr
# SLOWER than just numpy!
# """
# points = polyline[:,0]
# a = points[0:-2] # all "a" points
# b = points[1:-1] # b
# c = points[2:] # c points
# ax,ay = a[:,0],a[:,1]
# bx,by = b[:,0],b[:,1]
# cx,cy = c[:,0],c[:,1]
# # abx = '(bx - ax)'
# # aby = '(by - ay)'
# # cbx = '(bx - cx)'
# # cby = '(by - cy)'
# # # float dot = (ab.x * cb.x + ab.y * cb.y) dot product
# # dot = '%s * %s + %s * %s' %(abx,cbx,aby,cby)
# # # float cross = (ab.x * cb.y - ab.y * cb.x) cross product
# # cross = '(%s * %s - %s * %s)' %(abx,cby,aby,cbx)
# # # float alpha = atan2(cross, dot);
# # alpha = "arctan2(%s,%s)" %(cross,dot)
# # term = '%s*%s'%(alpha,180./np.pi)
# term = 'arctan2(((bx - ax) * (by - cy) - (by - ay) * (bx - cx)),(bx - ax) * (bx - cx) + (by - ay) * (by - cy))*57.2957795131'
# return ne.evaluate(term)
def split_at_angle(contour, curvature, angle):
"""
contour is array([[[108, 290]],[[111, 290]]], dtype=int32) shape=(number of points,1,dimension(2) )
curvature is a n-2 list
"""
segments = []
kink_index = [i for i in range(len(curvature)) if curvature[i] < angle]
for s,e in zip([0]+kink_index,kink_index+[None]): # list of slice indecies 0,i0,i1,i2,None
if e is not None:
segments.append(contour[s:e+1]) #need to include the last index
else:
segments.append(contour[s:e])
return segments
def find_kink(curvature, angle):
"""
contour is array([[[108, 290]],[[111, 290]]], dtype=int32) shape=(number of points,1,dimension(2) )
curvature is a n-2 list
"""
kinks = []
kink_index = [i for i in range(len(curvature)) if abs(curvature[i]) < angle]
return kink_index
def find_change_in_general_direction(curvature):
"""
return indecies of where the singn of curvature has flipped
"""
curv_pos = curvature > 0
split = []
currently_pos = curv_pos[0]
for c, is_pos in zip(range(curvature.shape[0]),curv_pos):
if is_pos !=currently_pos:
currently_pos = is_pos
split.append(c)
return split
def find_kink_and_dir_change(curvature,angle):
split = []
if curvature.shape[0] == 0:
return split
curv_pos = curvature > 0
currently_pos = curv_pos[0]
for idx,c, is_pos in zip(range(curvature.shape[0]),curvature,curv_pos):
if (is_pos !=currently_pos) or abs(c) < angle:
currently_pos = is_pos
split.append(idx)
return split
def find_slope_disc(curvature,angle = 15):
# this only makes sense when your polyline is longish
if len(curvature)<4:
return []
i = 2
split_idx = []
for anchor1,anchor2,candidate in zip(curvature,curvature[1:],curvature[2:]):
base_slope = anchor2-anchor1
new_slope = anchor2 - candidate
dif = abs(base_slope-new_slope)
if dif>=angle:
split_idx.add(i)
print i,dif
i +=1
return split_list
def find_slope_disc_test(curvature,angle = 15):
# this only makes sense when your polyline is longish
if len(curvature)<4:
return []
# mean = np.mean(curvature)
# print '------------------- start'
i = 2
split_idx = set()
for anchor1,anchor2,candidate in zip(curvature,curvature[1:],curvature[2:]):
base_slope = anchor2-anchor1
new_slope = anchor2 - candidate
dif = abs(base_slope-new_slope)
if dif>=angle:
split_idx.add(i)
# print i,dif
i +=1
i-= 3
for anchor1,anchor2,candidate in zip(curvature[::-1],curvature[:-1:][::-1],curvature[:-2:][::-1]):
avg = (anchor1+anchor2)/2.
dif = abs(avg-candidate)
if dif>=angle:
split_idx.add(i)
# print i,dif
i -=1
split_list = list(split_idx)
split_list.sort()
# print split_list
# print '-------end'
return split_list
def points_at_corner_index(contour,index):
"""
contour is array([[[108, 290]],[[111, 290]]], dtype=int32) shape=(number of points,1,dimension(2) )
#index n-2 because the curvature is n-2 (1st and last are not exsistent), this shifts the index (0 splits at first knot!)
"""
return [contour[i+1] for i in index]
def split_at_corner_index(contour,index):
"""
contour is array([[[108, 290]],[[111, 290]]], dtype=int32) shape=(number of points,1,dimension(2) )
#index n-2 because the curvature is n-2 (1st and last are not exsistent), this shifts the index (0 splits at first knot!)
"""
segments = []
index = [i+1 for i in index]
for s,e in zip([0]+index,index+[10000000]): # list of slice indecies 0,i0,i1,i2,
segments.append(contour[s:e+1])# +1 is for not loosing line segments
return segments
def convexity_defect(contour, curvature):
"""
contour is array([[[108, 290]],[[111, 290]]], dtype=int32) shape=(number of points,1,dimension(2) )
curvature is a n-2 list
"""
kinks = []
mean = np.mean(curvature)
if mean>0:
kink_index = [i for i in range(len(curvature)) if curvature[i] < 0]
else:
kink_index = [i for i in range(len(curvature)) if curvature[i] > 0]
for s in kink_index: # list of slice indecies 0,i0,i1,i2,None
kinks.append(contour[s+1]) # because the curvature is n-2 (1st and last are not exsistent)
return kinks,kink_index
def is_round(ellipse,ratio,tolerance=.8):
center, (axis1,axis2), angle = ellipse
if axis1 and axis2 and abs( ratio - min(axis2,axis1)/max(axis2,axis1)) < tolerance:
return True
else:
return False
def size_deviation(ellipse,target_size):
center, axis, angle = ellipse
return abs(target_size-max(axis))
def circle_grid(image, pattern_size=(4,11)):
"""Circle grid: finds an assymetric circle pattern
- circle_id: sorted from bottom left to top right (column first)
- If no circle_id is given, then the mean of circle positions is returned approx. center
- If no pattern is detected, function returns None
"""
status, centers = cv2.findCirclesGridDefault(image, pattern_size, flags=cv2.CALIB_CB_ASYMMETRIC_GRID)
if status:
return centers
else:
return None
def calibrate_camera(img_pts, obj_pts, img_size):
# generate pattern size
camera_matrix = np.zeros((3,3))
dist_coef = np.zeros(4)
rms, camera_matrix, dist_coefs, rvecs, tvecs = cv2.calibrateCamera(obj_pts, img_pts,
img_size, camera_matrix, dist_coef)
return camera_matrix, dist_coefs
def gen_pattern_grid(size=(4,11)):
pattern_grid = []
for i in xrange(size[1]):
for j in xrange(size[0]):
pattern_grid.append([(2*j)+i%2,i,0])
return np.asarray(pattern_grid, dtype='f4')
def normalize(pos, (width, height),flip_y=False):
"""
normalize return as float
"""
x = pos[0]
y = pos[1]
x /=float(width)
y /=float(height)
if flip_y:
return x,1-y
return x,y
def denormalize(pos, (width, height), flip_y=False):
"""
denormalize
"""
x = pos[0]
y = pos[1]
x *= width
if flip_y:
y = 1-y
y *= height
return x,y
def dist_pts_ellipse(((ex,ey),(dx,dy),angle),points):
"""
return unsigned euclidian distances of points to ellipse
"""
pts = np.float64(points)
rx,ry = dx/2., dy/2.
angle = (angle/180.)*np.pi
# ex,ey =ex+0.000000001,ey-0.000000001 #hack to make 0 divisions possible this is UGLY!!!
pts = pts - np.array((ex,ey)) # move pts to ellipse appears at origin , with this we copy data -deliberatly!
M_rot = np.mat([[np.cos(angle),-np.sin(angle)],[np.sin(angle),np.cos(angle)]])
pts = np.array(pts*M_rot) #rotate so that ellipse axis align with coordinate system
# print "rotated",pts
pts /= np.array((rx,ry)) #normalize such that ellipse radii=1
# print "normalize",norm_pts
norm_mag = np.sqrt((pts*pts).sum(axis=1))
norm_dist = abs(norm_mag-1) #distance of pt to ellipse in scaled space
# print 'norm_mag',norm_mag
# print 'norm_dist',norm_dist
ratio = (norm_dist)/norm_mag #scale factor to make the pts represent their dist to ellipse
# print 'ratio',ratio
scaled_error = np.transpose(pts.T*ratio) # per vector scalar multiplication: makeing sure that boradcasting is done right
# print "scaled error points", scaled_error
real_error = scaled_error*np.array((rx,ry))
# print "real point",real_error
error_mag = np.sqrt((real_error*real_error).sum(axis=1))
# print 'real_error',error_mag
# print 'result:',error_mag
return error_mag
if ne:
def dist_pts_ellipse(((ex,ey),(dx,dy),angle),points):
"""
return unsigned euclidian distances of points to ellipse
same as above but uses numexpr for 2x speedup
"""
pts = np.float64(points)
pts.shape=(-1,2)
rx,ry = dx/2., dy/2.
angle = (angle/180.)*np.pi
# ex,ey = ex+0.000000001 , ey-0.000000001 #hack to make 0 divisions possible this is UGLY!!!
x = pts[:,0]
y = pts[:,1]
# px = '((x-ex) * cos(angle) + (y-ey) * sin(angle))/rx'
# py = '(-(x-ex) * sin(angle) + (y-ey) * cos(angle))/ry'
# norm_mag = 'sqrt(('+px+')**2+('+py+')**2)'
# norm_dist = 'abs('+norm_mag+'-1)'
# ratio = norm_dist + "/" + norm_mag
# x_err = ''+px+'*'+ratio+'*rx'
# y_err = ''+py+'*'+ratio+'*ry'
# term = 'sqrt(('+x_err+')**2 + ('+y_err+')**2 )'
term = 'sqrt((((x-ex) * cos(angle) + (y-ey) * sin(angle))/rx*abs(sqrt((((x-ex) * cos(angle) + (y-ey) * sin(angle))/rx)**2+((-(x-ex) * sin(angle) + (y-ey) * cos(angle))/ry)**2)-1)/sqrt((((x-ex) * cos(angle) + (y-ey) * sin(angle))/rx)**2+((-(x-ex) * sin(angle) + (y-ey) * cos(angle))/ry)**2)*rx)**2 + ((-(x-ex) * sin(angle) + (y-ey) * cos(angle))/ry*abs(sqrt((((x-ex) * cos(angle) + (y-ey) * sin(angle))/rx)**2+((-(x-ex) * sin(angle) + (y-ey) * cos(angle))/ry)**2)-1)/sqrt((((x-ex) * cos(angle) + (y-ey) * sin(angle))/rx)**2+((-(x-ex) * sin(angle) + (y-ey) * cos(angle))/ry)**2)*ry)**2 )'
error_mag = ne.evaluate(term)
return error_mag
def metric(l):
"""
example metric for search
"""
# print 'evaluating', idecies
global evals
evals +=1
return sum(l) < 3
def pruning_quick_combine(l,fn,seed_idx=None,max_evals=1e20,max_depth=5):
"""
l is a list of object to quick_combine.
the evaluation fn should accept idecies to your list and the list
it should return a binary result on wether this set is good
this search finds all combinations but assumes:
that a bad subset can not be bettered by adding more nodes
that a good set may not always be improved by a 'passing' superset (purging subsets will revoke this)
if all items and their combinations pass the evaluation fn you get n**2 -1 solutions
which leads to (2**n - 1) calls of your evaluation fn
it needs more evaluations than finding strongly connected components in a graph because:
(1,5) and (1,6) and (5,6) may work but (1,5,6) may not pass evaluation, (n,m) being list idx's
"""
if seed_idx:
non_seed_idx = [i for i in range(len(l)) if i not in seed_idx]
else:
#start from every item
seed_idx = range(len(l))
non_seed_idx = []
mapping = seed_idx+non_seed_idx
unknown = [[node] for node in range(len(seed_idx))]
# print mapping
results = []
prune = []
while unknown and max_evals:
path = unknown.pop(0)
max_evals -= 1
# print '@idx',[mapping[i] for i in path]
# print '@content',path
if not len(path) > max_depth:
# is this combination even viable, or did a subset fail already?
if not any(m.issubset(set(path)) for m in prune):
#we have not tested this and a subset of this was sucessfull before
if fn([l[mapping[i]] for i in path]):
# yes this was good, keep as solution
results.append([mapping[i] for i in path])
# lets explore more by creating paths to each remaining node
decedents = [path+[i] for i in range(path[-1]+1,len(mapping)) ]
unknown.extend(decedents)
else:
# print "pruning",path
prune.append(set(path))
return results
# def is_subset(needle,haystack):
# """ Check if needle is ordered subset of haystack in O(n)
# taken from:
# http://stackoverflow.com/questions/1318935/python-list-filtering-remove-subsets-from-list-of-lists
# """
# if len(haystack) < len(needle): return False
# index = 0
# for element in needle:
# try:
# index = haystack.index(element, index) + 1
# except ValueError:
# return False
# else:
# return True
# def filter_subsets(lists):
# """ Given list of lists, return new list of lists without subsets
# taken from:
# http://stackoverflow.com/questions/1318935/python-list-filtering-remove-subsets-from-list-of-lists
# """
# for needle in lists:
# if not any(is_subset(needle, haystack) for haystack in lists
# if needle is not haystack):
# yield needle
def filter_subsets(l):
return [m for i, m in enumerate(l) if not any(set(m).issubset(set(n)) for n in (l[:i] + l[i+1:]))]
if __name__ == '__main__':
# tst = []
# for x in range(10):
# tst.append(gen_pattern_grid())
# tst = np.asarray(tst)
# print tst.shape
#test polyline
# *-* *
# | \ |
# * *-*
# |
# *-*
pl = np.array([[[0, 0]],[[0, 1]],[[1, 1]],[[2, 1]],[[2, 2]],[[1, 3]],[[1, 4]],[[2,4]]], dtype=np.int32)
curvature = GetAnglesPolyline(pl,closed=0)
print curvature
curvature = GetAnglesPolyline(pl,closed=1)
# print curvature
# print find_curv_disc(curvature)
# idx = find_kink_and_dir_change(curvature,60)
# print idx
# print split_at_corner_index(pl,idx)
# ellipse = ((0,0),(np.sqrt(2),np.sqrt(2)),0)
# pts = np.array([(0,1),(.5,.5),(0,-1)])
# # print pts.dtype
# print dist_pts_ellipse(ellipse,pts)
# print pts
# # print test()
# l = [1,2,1,0,1,0]
# print len(l)
# # evals = 0
# # r = quick_combine(l,metric)
# # # print r
# # print filter_subsets(r)
# # print evals
# evals = 0
# r = pruning_quick_combine(l,metric,[2])
# print r
# print filter_subsets(r)
# print evals