better grid optimization

[imago.git] / gridf.py

diff --git a/gridf.py b/gridf.py
index 17be372..6f044f1 100644 (file)
--- a/gridf.py
+++ b/gridf.py
@@ -1,8 +1,12 @@
+import multiprocessing
+

 import Image, ImageDraw, ImageFilter
 
 import Image, ImageDraw, ImageFilter
 

-from manual import lines as g_grid, l2ad
+from geometry import V
+from manual import lines as g_grid, l2ad, intersection, line as g_line

 from intrsc import intersections_from_angl_dist
 from linef import line_from_angl_dist
 from intrsc import intersections_from_angl_dist
 from linef import line_from_angl_dist

+import pcf

 
 class GridFittingFailedError(Exception):
     pass
 
 class GridFittingFailedError(Exception):
     pass


@@ -15,50 +19,117 @@ class MyGaussianBlur(ImageFilter.Filter):
     def filter(self, image):
         return image.gaussian_blur(self.radius)
 
     def filter(self, image):
         return image.gaussian_blur(self.radius)
 

-class V():
-    def __init__(self, x, y):
-        self.x = x
-        self.y = y
-    
-    def __add__(self, other):
-        return V(self.x + other.x, self.y + other.y)
-
-    def __sub__(self, other):
-        return V(self.x - other.x, self.y - other.y)
-
-    def __rmul__(self, other):
-        return V(other * self.x, other * self.y)
-
-    def t(self):
-        return (self.x, self.y)
-
-def find(lines, size, l1, l2, bounds, hough, do_something):
+def projection(point, line, vector):
+    return V(*intersection(g_line(point, point + vector.normal), g_line(*line)))
+
+def job_br1(args):
+    X, Y, im_l, a, b, c, d, s, v1, v2, k, hough, size = args
+    return [(distance(im_l, 
+                     get_grid(a + X[y] * s * v1, 
+                              b + Y[y] * s * v1, 
+                              c, d, hough, size),
+                     size), a + X[y] * s * v1, b + Y[y] * s * v1) for y in range(2 *k)]
+
+def job_br2(args):
+    X, Y, im_l, a, b, c, d, s, v1, v2, k, hough, size = args
+    return [(distance(im_l, 
+                     get_grid(a, b, c + X[y] * s * v2, 
+                              d + Y[y] * s * v2, 
+                              hough, size),
+                     size), c + X[y] * s * v2, d + Y[y] * s * v2) for y in range(2 *k)]
+
+def find(lines, size, l1, l2, bounds, hough, do_something, im_h):

     a, b, c, d = [V(*a) for a in bounds]
     l1 = line_from_angl_dist(l1, size)
     l2 = line_from_angl_dist(l2, size)
     v1 = V(*l1[0]) - V(*l1[1])
     v2 = V(*l2[0]) - V(*l2[1])
     a, b, c, d = [V(*a) for a in bounds]
     l1 = line_from_angl_dist(l1, size)
     l2 = line_from_angl_dist(l2, size)
     v1 = V(*l1[0]) - V(*l1[1])
     v2 = V(*l2[0]) - V(*l2[1])

+    a = projection(a, l1, v1) 
+    b = projection(b, l1, v1) 
+    c = projection(c, l2, v2) 
+    d = projection(d, l2, v2) 
+
+    im_l = Image.new('L', size)
+    dr_l = ImageDraw.Draw(im_l)
+    for line in sum(lines, []):
+        dr_l.line(line_from_angl_dist(line, size), width=1, fill=255)
+    #im_l = im_l.filter(MyGaussianBlur(radius=30))
+    #GaussianBlur is undocumented class, may not work in future versions of PIL
+    #im_l = im_l.tostring()
+    im_l = im_h.tostring() # hocus pocus
+
+    #from gridf_analyzer import error_surface
+    #error_surface(im_l, a, b, c, d, hough, size, v1 ,v2)
+

     grid = get_grid(a, b, c, d, hough, size)
     grid = get_grid(a, b, c, d, hough, size)

-    dist = distance(lines, grid, size)
-    print dist
+    dist = distance(im_l, grid, size)

     
     

-    s = 0.02
-    while True:
-        ts1 = [(s, 0), (-s, 0), (s, s), (-s, -s), (-s, s), (s, -s), (0, s),  (0, -s)]
+    #let's try the bruteforce aproach:
+    s = 0.001
+    k = 50
+    X, Y = [], []
+    for i in range(-k, k):
+        X.append(range(-k, k))
+        Y.append(2*k*[i])
+
+    pool = multiprocessing.Pool(None)
+
+    tasks = [(X[x], Y[x], im_l, a, b, c, d, s, v1, v2, k, hough, size) for x in xrange(0, 2 * k)]
+
+    #start = time.time()
+    opt_ab = pool.map(job_br1, tasks, 1)
+    opt_cd = pool.map(job_br2, tasks, 1)
+    an, bn, cn, dn = 4 * [0]
+    d1 = 0
+    for lst in opt_ab:
+        for tpl in lst:
+            if tpl[0] > d1:
+                d1 = tpl[0]
+                a, b = tpl[1], tpl[2]
+    d1 = 0
+    for lst in opt_cd:
+        for tpl in lst:
+            if tpl[0] > d1:
+                d1 = tpl[0]
+                c, d = tpl[1], tpl[2]
+    #print time.time() - start
+    grid = get_grid(a, b, c, d, hough, size) 
+    grid_lines = [[l2ad(l, size) for l in grid[0]], [l2ad(l, size) for l in grid[1]]]
+    return grid, grid_lines
+
+    #old optimization experiments: 
+    print dist
+   
+    path = [(0,0)] #MNTR
+    s = 0.01
+    for _ in range(10):
+        ts1 = [(s, 0), (0, s), (-s, 0), (0, -s)]

         grids = [(get_grid(a + t[0] * v1, b + t[1] * v1, 
                            c, d, hough, size), t) for t in ts1]
         grids = [(get_grid(a + t[0] * v1, b + t[1] * v1, 
                            c, d, hough, size), t) for t in ts1]

-        distances = [(distance(lines, grid, size), 
-                      grid, t) for grid, t in grids]
-        distances.sort(reverse=True)
-        if distances[0][0] > dist:
-            dist = distances[0][0]
-            grid = distances[0][1]
-            t = distances[0][2]
-            a, b = a + t[0] * v1, b + t[1] * v1
-            print dist
-            s *= 0.75
-        else: 
-           break
+        distances = [distance(im_l, grid, size) for (grid, t) in grids]
+        gradient = [(di - dist) for di in distances]
+        gradient = [gradient[0] - gradient[2], gradient[1] - gradient[3]]
+        norm = (gradient[0] ** 2 + gradient[1] ** 2) ** 0.5
+        gradient = [g / (100 * norm) for g in gradient]
+        path.append(gradient)
+        a, b = a + gradient[0] * v1, b + gradient[1] * v1
+        dist = distance(im_l, grid, size)
+        print dist
+
+    ###MNTR
+    import matplotlib.pyplot as plt
+    from matplotlib import cm
+    import pickle
+
+    X, Y, Z = pickle.load(open('surface250'))
+
+    plt.imshow(Z, cmap=cm.jet, interpolation='none', 
+               origin='upper', extent=(-0.250, 0.250, -0.250, 0.250), aspect='equal')
+    plt.colorbar()
+    plt.plot([y for (x, y) in path], [x for (x, y) in path], 'go-')
+
+    plt.show()
+    ###MNTR

 
     print "---"
 
 
     print "---"
 


@@ -67,7 +138,7 @@ def find(lines, size, l1, l2, bounds, hough, do_something):
         ts1 = [(s, 0), (-s, 0), (s, s), (-s, -s), (-s, s), (s, -s), (0, s),  (0, -s)]
         grids = [(get_grid(a, b, 
                            c + t[0] * v2, d + t[1] * v2, hough, size), t) for t in ts1]
         ts1 = [(s, 0), (-s, 0), (s, s), (-s, -s), (-s, s), (s, -s), (0, s),  (0, -s)]
         grids = [(get_grid(a, b, 
                            c + t[0] * v2, d + t[1] * v2, hough, size), t) for t in ts1]

-        distances = [(distance(lines, grid, size), 
+        distances = [(distance(im_l, grid, size), 

                       grid, t) for grid, t in grids]
         distances.sort(reverse=True)
         if distances[0][0] > dist:
                       grid, t) for grid, t in grids]
         distances.sort(reverse=True)
         if distances[0][0] > dist:


@@ -84,9 +155,10 @@ def find(lines, size, l1, l2, bounds, hough, do_something):
     return grid, grid_lines
 
 def get_grid(a, b, c, d, hough, size):
     return grid, grid_lines
 
 def get_grid(a, b, c, d, hough, size):

-    l1 = hough.lines_from_list([a.t(), b.t()])
-    l2 = hough.lines_from_list([c.t(), d.t()])
+    l1 = hough.lines_from_list([a, b])
+    l2 = hough.lines_from_list([c, d])

     c = intersections_from_angl_dist([l1, l2], size, get_all=True)
     c = intersections_from_angl_dist([l1, l2], size, get_all=True)

+    #TODO do something when a corner is outside the image

     corners = (c[0] + c[1])
     if len(corners) < 4:
         print l1, l2, c
     corners = (c[0] + c[1])
     if len(corners) < 4:
         print l1, l2, c


@@ -94,25 +166,31 @@ def get_grid(a, b, c, d, hough, size):
     grid = g_grid(corners)
     return grid
 
     grid = g_grid(corners)
     return grid
 

-def distance(lines, grid, size):
-    im_l = Image.new('L', size)
-    dr_l = ImageDraw.Draw(im_l)
-    for line in sum(lines, []):
-        dr_l.line(line_from_angl_dist(line, size), width=1, fill=255)
-    im_l = im_l.filter(MyGaussianBlur(radius=3))
-    # GaussianBlur is undocumented class, may not work in future versions of PIL
+def line_out(line, size):
+    for p in line:
+        if p[0] < 0 or p[0] > size[0] or p[1] < 0 or p[1] > size[1]:
+            return True
+    else:
+        return False
+
+def distance(im_l, grid, size):

     im_g = Image.new('L', size)
     dr_g = ImageDraw.Draw(im_g)
     for line in grid[0] + grid[1]:
         dr_g.line(line, width=1, fill=255)
     im_g = Image.new('L', size)
     dr_g = ImageDraw.Draw(im_g)
     for line in grid[0] + grid[1]:
         dr_g.line(line, width=1, fill=255)

-    im_d, distance = combine(im_l, im_g)
+        if line_out(line, size):
+            return 0
+    #im_g = im_g.filter(MyGaussianBlur(radius=3))
+    #GaussianBlur is undocumented class, may not work in future versions of PIL
+    #im_d, distance = combine(im_l, im_g)
+    distance = pcf.combine(im_l, im_g.tostring())

     return distance
 
 def combine(bg, fg):
     bg_l = bg.load()
     fg_l = fg.load()
     return distance
 
 def combine(bg, fg):
     bg_l = bg.load()
     fg_l = fg.load()

-    res = Image.new('L', fg.size)
-    res_l = res.load()
+    #res = Image.new('L', fg.size)
+    #res_l = res.load()

 
     score = 0
     area = 0
 
     score = 0
     area = 0


@@ -120,8 +198,9 @@ def combine(bg, fg):
     for x in xrange(fg.size[0]):
         for y in xrange(fg.size[1]):
             if fg_l[x, y]:
     for x in xrange(fg.size[0]):
         for y in xrange(fg.size[1]):
             if fg_l[x, y]:

-                res_l[x, y] = bg_l[x, y]
+                #res_l[x, y] = bg_l[x, y] * fg_l[x, y]

                 score +=  bg_l[x, y]
                 area += 1
 
                 score +=  bg_l[x, y]
                 area += 1
 

-    return res, float(score)/area
+    #return res, float(score)/area
+    return None, float(score)/area