bgs/local_svd_2016.py at master · machinelearningblr/bgs · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
import numpy as np
import cv2
import argparse
import os
import time
import random
import maxflow

argparser = argparse.ArgumentParser(description='Draft implementation of the VIBE background subtraction algorithm.')

argparser.add_argument('-g', '--gt', help='Directory with ground-truth frames', required=True)
argparser.add_argument('-f', '--frames', help='Directory with input frames', required=True)
argparser.add_argument('-k', '--output', type=int, help='Calculate answer for K-th frame and output')
args = argparser.parse_args()

gt = map(lambda x: os.path.join(args.gt, x), os.listdir(args.gt))
gt.sort()
f = map(lambda x: os.path.join(args.frames, x), os.listdir(args.frames))
f.sort()

gt = np.float32(map(lambda x: cv2.imread(x, cv2.IMREAD_GRAYSCALE), gt)) / 255.0
f = np.float32(map(lambda x: cv2.imread(x, cv2.IMREAD_COLOR), f)) / 255.0

S = 20
H, W = f[0].shape[:2]
samples_int = np.zeros((S,) + f[0].shape, np.float32)
samples_lsbp = np.zeros((S,H,W,9), np.bool)
D = np.ones((S,H,W), np.float32) * 0.2
Racc = np.ones((H,W), np.float32) * 0.2
Tacc = np.ones((H,W), np.float32) * 0.08

def extract_LSBP(frame, R1=1, R2=1, tau=0.05):
	intensity = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
	intens = np.zeros((H+2*R1, W+2*R1), np.float32)
	intens[R1:H+R1, R1:W+R1] = intensity
	g = np.zeros((H+2*R2, W+2*R2), np.float32)
	for i in xrange(H):
		for j in xrange(W):
			s = np.linalg.svd(intens[i:i+2*R1+1,j:j+2*R1+1], compute_uv=False)
			g[i+R2,j+R2] = (s[1] + s[2]) / s[0]
	lsbp = np.zeros((H, W, (2*R2+1)**2), np.bool)
	for i in xrange(H):
		for j in xrange(W):
			lsbp[i,j] = (np.abs(g[i:i+2*R2+1,j:j+2*R2+1] - g[i+R2,j+R2]) < tau).ravel()
	return lsbp

def SVD_init(frame, R=S/2):
	lsbp = extract_LSBP(frame)
	for i in xrange(H):
		for j in xrange(W):
			i0 = np.clip(i, R, H-R-1)
			j0 = np.clip(j, R, W-R-1)
			samples_int[0,i,j] = frame[i,j]
			samples_lsbp[0,i,j] = lsbp[i,j]
			for k in xrange(1, R):
				i1, j1 = i0+random.randrange(-R, R+1), j0+random.randrange(-R, R+1)
				samples_int[k,i,j] = frame[i1,j1]
				samples_lsbp[k,i,j] = lsbp[i1,j1]

SVD_init(f[0])

def SVD_step(frame, HR=4, threshold=2, Rscale=5, Rlr=0.1, Tlr=0.02):
	global Racc
	global Tacc
	lsbp = extract_LSBP(frame)
	Racc = (1.0 - Rlr) * Racc + Rlr * (np.mean(D, axis=0) * Rscale)
	mask = np.float32(np.sum((np.sum(np.abs(samples_int - frame), axis=3) < Racc) & (np.sum(samples_lsbp ^ lsbp, axis=3) < HR), axis=0) < threshold)
	Tacc = (1.0 - Tlr) * Tacc + Tlr * (1.0 - mask)
	Tacc = np.clip(Tacc, 0.05, 0.8)
	for i in xrange(H):
		for j in xrange(W):
			if mask[i,j] == 0.0:
				if random.random() < Tacc[i,j]:
					k = random.randrange(0, S)
					dist = np.sum(np.abs(samples_int[:,i,j] - frame[i,j]), axis=1)
					dist.sort()
					D[k,i,j] = dist[1]
					samples_int[k,i,j] = frame[i,j]
					samples_lsbp[k,i,j] = lsbp[i,j]
				if random.random() < Tacc[i,j]:
					k = random.randrange(0, S)
					i0, j0 = np.clip(i+random.randrange(-1,2), 0, H-1), np.clip(j+random.randrange(-1,2), 0, W-1)
					dist = np.sum(np.abs(samples_int[:,i0,j0] - frame[i,j]), axis=1)
					dist.sort()
					D[k,i0,j0] = dist[1]
					samples_int[k,i0,j0] = frame[i,j]
					samples_lsbp[k,i0,j0] = lsbp[i,j]
	return mask

def postprocessing(im, unary):
	unary = np.float32(unary)
	unary = cv2.GaussianBlur(unary, (9, 9), 0)

	g = maxflow.Graph[float]()
	nodes = g.add_grid_nodes(unary.shape)

	for i in xrange(im.shape[0]):
		for j in xrange(im.shape[1]):
			v = nodes[i,j]
			g.add_tedge(v, -unary[i,j], -1.0+unary[i,j])

	def potts_add_edge(i0, j0, i1, j1):
		v0, v1 = nodes[i0,j0], nodes[i1,j1]
		w = 0.1 * np.exp(-((im[i0,j0] - im[i1,j1])**2).sum() / 0.1)
		g.add_edge(v0, v1, w, w)

	for i in xrange(1,im.shape[0]-1):
		for j in xrange(1,im.shape[1]-1):
			potts_add_edge(i, j, i, j-1)
			potts_add_edge(i, j, i, j+1)
			potts_add_edge(i, j, i-1, j)
			potts_add_edge(i, j, i+1, j)

	g.maxflow()
	seg = np.float32(g.get_grid_segments(nodes))
	return seg

if args.output is not None:
	for i in xrange(f.shape[0]):
		sec = time.time()
		out = SVD_step(f[i])
		mrf = postprocessing(f[i], out)
		print('Frame %d, %.3f sec.' % (i, time.time() - sec))
		if i >= args.output:
			cv2.imwrite('svd-frame.png', f[i] * 255)
			cv2.imwrite('svd-mask.png', out * 255)
			cv2.imwrite('svd-gt.png', gt[i] * 255)
			cv2.imwrite('svd-mrf.png', mrf * 255)
			break
else:
	for i in xrange(f.shape[0]):
		cv2.imshow('Frame', f[i])
		cv2.imshow('Ground-truth', gt[i])
		out = SVD_step(f[i])
		cv2.imshow('Output', out)
		cv2.imshow('Output + MRF', postprocessing(f[i], out))
		k = cv2.waitKey(0)
		if k == 27:
			break