Skip to content

Commit 236b13d

Browse files
davidsandbergdavidsandberg
committed
First version of MTCNN face detection and alignment
1 parent 1c9024f commit 236b13d

14 files changed

+781
-0
lines changed

Diff for: ‎__init__.py

Whitespace-only changes.

Diff for: ‎src/align/__init__.py

Whitespace-only changes.

Diff for: ‎src/align/__init__.pyc

139 Bytes
Binary file not shown.

Diff for: ‎src/align/det1.py

+23
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,23 @@
1+
#from __future__ import absolute_import
2+
from __future__ import division
3+
from __future__ import print_function
4+
5+
from src.align.network import Network
6+
#import Network
7+
8+
class PNet(Network):
9+
def setup(self):
10+
(self.feed('data')
11+
.conv(3, 3, 10, 1, 1, padding='VALID', relu=False, name='conv1')
12+
.prelu(name='PReLU1')
13+
.max_pool(2, 2, 2, 2, name='pool1')
14+
.conv(3, 3, 16, 1, 1, padding='VALID', relu=False, name='conv2')
15+
.prelu(name='PReLU2')
16+
.conv(3, 3, 32, 1, 1, padding='VALID', relu=False, name='conv3')
17+
.prelu(name='PReLU3')
18+
.conv(1, 1, 2, 1, 1, relu=False, name='conv4-1')
19+
.softmax(3,name='prob1'))
20+
21+
(self.feed('PReLU3')
22+
.conv(1, 1, 4, 1, 1, relu=False, name='conv4-2'))
23+

Diff for: ‎src/align/det1.pyc

1.17 KB
Binary file not shown.

Diff for: ‎src/align/det2.py

+24
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,24 @@
1+
from __future__ import absolute_import
2+
from __future__ import division
3+
from __future__ import print_function
4+
5+
from network import Network
6+
7+
class RNet(Network):
8+
def setup(self):
9+
(self.feed('data')
10+
.conv(3, 3, 28, 1, 1, padding='VALID', relu=False, name='conv1')
11+
.prelu(name='prelu1')
12+
.max_pool(3, 3, 2, 2, name='pool1')
13+
.conv(3, 3, 48, 1, 1, padding='VALID', relu=False, name='conv2')
14+
.prelu(name='prelu2')
15+
.max_pool(3, 3, 2, 2, padding='VALID', name='pool2')
16+
.conv(2, 2, 64, 1, 1, padding='VALID', relu=False, name='conv3')
17+
.prelu(name='prelu3')
18+
.fc(128, relu=False, name='conv4')
19+
.prelu(name='prelu4')
20+
.fc(2, relu=False, name='conv5-1')
21+
.softmax(1,name='prob1'))
22+
23+
(self.feed('prelu4')
24+
.fc(4, relu=False, name='conv5-2'))

Diff for: ‎src/align/det2.pyc

1.3 KB
Binary file not shown.

Diff for: ‎src/align/det3.py

+30
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,30 @@
1+
from __future__ import absolute_import
2+
from __future__ import division
3+
from __future__ import print_function
4+
5+
from network import Network
6+
7+
class ONet(Network):
8+
def setup(self):
9+
(self.feed('data')
10+
.conv(3, 3, 32, 1, 1, padding='VALID', relu=False, name='conv1')
11+
.prelu(name='prelu1')
12+
.max_pool(3, 3, 2, 2, name='pool1')
13+
.conv(3, 3, 64, 1, 1, padding='VALID', relu=False, name='conv2')
14+
.prelu(name='prelu2')
15+
.max_pool(3, 3, 2, 2, padding='VALID', name='pool2')
16+
.conv(3, 3, 64, 1, 1, padding='VALID', relu=False, name='conv3')
17+
.prelu(name='prelu3')
18+
.max_pool(2, 2, 2, 2, name='pool3')
19+
.conv(2, 2, 128, 1, 1, padding='VALID', relu=False, name='conv4')
20+
.prelu(name='prelu4')
21+
.fc(256, relu=False, name='conv5')
22+
.prelu(name='prelu5')
23+
.fc(2, relu=False, name='conv6-1')
24+
.softmax(1, name='prob1'))
25+
26+
(self.feed('prelu5')
27+
.fc(4, relu=False, name='conv6-2'))
28+
29+
(self.feed('prelu5')
30+
.fc(10, relu=False, name='conv6-3'))

Diff for: ‎src/align/det3.pyc

1.46 KB
Binary file not shown.

Diff for: ‎src/align/detect_face.py

+293
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,293 @@
1+
""" Hej
2+
"""
3+
import numpy as np
4+
import cv2
5+
from math import floor
6+
#from scipy import misc
7+
import scipy.io as sio
8+
9+
def detect_face(img, minsize, pnet, rnet, onet, threshold, factor):
10+
# im: input image
11+
# minsize: minimum of faces' size
12+
# pnet, rnet, onet: caffemodel
13+
# threshold: threshold=[th1 th2 th3], th1-3 are three steps's threshold
14+
# fastresize: resize img from last scale (using in high-resolution images) if fastresize==true
15+
factor_count=0
16+
total_boxes=np.empty((0,9))
17+
points=[]
18+
h=img.shape[0]
19+
w=img.shape[1]
20+
minl=np.amin([h, w])
21+
#img=single(img);
22+
# if fastresize
23+
# im_data=(single(img)-127.5)*0.0078125;
24+
# end
25+
m=12.0/minsize
26+
minl=minl*m
27+
# creat scale pyramid
28+
scales=[]
29+
while minl>=12:
30+
scales += [m*np.power(factor, factor_count)]
31+
minl = minl*factor
32+
factor_count += 1
33+
34+
# first stage
35+
#for j = 1:size(scales,2)
36+
for j in range(len(scales)):
37+
scale=scales[j]
38+
hs=int(np.ceil(h*scale))
39+
ws=int(np.ceil(w*scale))
40+
#im_data=(imResample(img,[hs ws],'bilinear')-127.5)*0.0078125;
41+
#im_data = (misc.imresize(img, (hs, ws), interp='bilinear')-127.5)*0.0078125
42+
#im_data = (cv2.resize(img, (hs, ws), interpolation=cv2.INTER_LINEAR)-127.5)*0.0078125
43+
#im_data = (cv2.resize(img, (hs, ws), interpolation=cv2.INTER_NEAREST)-127.5)*0.0078125
44+
#im_data = (img[0:hs,0:ws,:]-127.5)*0.0078125
45+
im_data = imResample2(img, (hs, ws), 'bilinear')
46+
im_data = (im_data-127.5)*0.0078125
47+
48+
# PNet.blobs('data').reshape([hs ws 3 1]);
49+
# out=PNet.forward({im_data});
50+
51+
img_x = np.expand_dims(im_data, 0)
52+
img_y = np.transpose(img_x, (0,2,1,3))
53+
out = pnet(img_y)
54+
out0 = np.transpose(out[0], (0,2,1,3))
55+
#out0 = out[0]
56+
out1 = np.transpose(out[1], (0,2,1,3))
57+
#out1 = out[1]
58+
59+
#boxes=generateBoundingBox(out{2}(:,:,2), out{1}, scale, threshold[0]);
60+
boxes, reg = generateBoundingBox(out1[0,:,:,1].copy(), out0[0,:,:,:].copy(), scale, threshold[0])
61+
62+
# inter-scale nms
63+
pick = nms(boxes.copy(), 0.5, 'Union')
64+
boxes = boxes[pick,:] # boxes=boxes(pick,:);
65+
if boxes.size>0: # if ~isempty(boxes)
66+
total_boxes = np.append(total_boxes, boxes, axis=0) # total_boxes=[total_boxes;boxes];
67+
68+
numbox = total_boxes.shape[0] # numbox=size(total_boxes,1);
69+
if numbox>0: # if ~isempty(total_boxes)
70+
pick = nms(total_boxes.copy(), 0.7, 'Union') # pick=nms(total_boxes,0.7,'Union');
71+
total_boxes = total_boxes[pick,:] # total_boxes=total_boxes(pick,:);
72+
regw = total_boxes[:,2]-total_boxes[:,0] # regw=total_boxes(:,3)-total_boxes(:,1);
73+
regh = total_boxes[:,3]-total_boxes[:,1] # regh=total_boxes(:,4)-total_boxes(:,2);
74+
# total_boxes=[total_boxes(:,1)+total_boxes(:,6).*regw total_boxes(:,2)+total_boxes(:,7).*regh total_boxes(:,3)+total_boxes(:,8).*regw total_boxes(:,4)+total_boxes(:,9).*regh total_boxes(:,5)];
75+
qq1 = total_boxes[:,0]+total_boxes[:,5]*regw
76+
qq2 = total_boxes[:,1]+total_boxes[:,6]*regh
77+
qq3 = total_boxes[:,2]+total_boxes[:,7]*regw
78+
qq4 = total_boxes[:,3]+total_boxes[:,8]*regh
79+
total_boxes = np.transpose(np.vstack([qq1, qq2, qq3, qq4, total_boxes[:,4]]))
80+
total_boxes = rerec(total_boxes.copy()) # total_boxes=rerec(total_boxes);
81+
total_boxes[:,0:4] = np.fix(total_boxes[:,0:4]) # total_boxes(:,1:4)=fix(total_boxes(:,1:4));
82+
dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph = pad(total_boxes.copy(), w, h)
83+
84+
numbox = total_boxes.shape[0] # numbox=size(total_boxes,1);
85+
if numbox>0:
86+
# second stage
87+
tempimg = np.zeros((24,24,3,numbox)) # tempimg=zeros(24,24,3,numbox);
88+
for k in range(0,numbox): # for k=1:numbox
89+
tmp = np.zeros((tmph[k],tmpw[k],3)) # tmp=zeros(tmph(k),tmpw(k),3);
90+
tmp[dy[k]-1:edy[k],dx[k]-1:edx[k],:] = img[y[k]-1:ey[k],x[k]-1:ex[k],:] # tmp(dy(k):edy(k),dx(k):edx(k),:)=img(y(k):ey(k),x(k):ex(k),:);
91+
if tmp.shape[0]>0 and tmp.shape[1]>0 or tmp.shape[0]==0 and tmp.shape[1]==0: # if size(tmp,1)>0 && size(tmp,2)>0 || size(tmp,1)==0 && size(tmp,2)==0
92+
tempimg[:,:,:,k] = imResample2(tmp, (24, 24), 'bilinear') # tempimg(:,:,:,k)=imResample(tmp,[24 24],'bilinear');
93+
else:
94+
return np.empty() # total_boxes = []; return;
95+
tempimg = (tempimg-127.5)*0.0078125 # tempimg=(tempimg-127.5)*0.0078125;
96+
# RNet.blobs('data').reshape([24 24 3 numbox]);
97+
# out=RNet.forward({tempimg});
98+
tempimg1 = np.transpose(tempimg, (3,1,0,2))
99+
out = rnet(tempimg1)
100+
out0 = np.transpose(out[0])
101+
out1 = np.transpose(out[1])
102+
score = out1[1,:] # score=squeeze(out{2}(2,:));
103+
ipass = np.where(score>threshold[1]) # pass=find(score>threshold(2));
104+
total_boxes = np.hstack([total_boxes[ipass[0],0:4].copy(), np.expand_dims(score[ipass].copy(),1)]) # total_boxes=[total_boxes(pass,1:4) score(pass)'];
105+
mv = out0[:,ipass[0]] # mv=out{1}(:,pass);
106+
if total_boxes.shape[0]>0: # if size(total_boxes,1)>0
107+
pick = nms(total_boxes, 0.7, 'Union') # pick=nms(total_boxes,0.7,'Union');
108+
total_boxes = total_boxes[pick,:] # total_boxes=total_boxes(pick,:);
109+
total_boxes = bbreg(total_boxes.copy(), np.transpose(mv[:,pick])) # total_boxes=bbreg(total_boxes,mv(:,pick)');
110+
total_boxes = rerec(total_boxes.copy()) # total_boxes=rerec(total_boxes);
111+
112+
numbox = total_boxes.shape[0] # numbox=size(total_boxes,1);
113+
if numbox>0:
114+
# third stage
115+
total_boxes=np.fix(total_boxes) # total_boxes=fix(total_boxes);
116+
dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph = pad(total_boxes.copy(), w, h) # [dy edy dx edx y ey x ex tmpw tmph]=pad(total_boxes,w,h);
117+
tempimg = np.zeros((48,48,3,numbox)) # tempimg=zeros(48,48,3,numbox);
118+
for k in range(0,numbox): # for k=1:numbox
119+
tmp = np.zeros((tmph[k],tmpw[k],3))# tmp=zeros(tmph(k),tmpw(k),3);
120+
tmp[dy[k]-1:edy[k],dx[k]-1:edx[k],:] = img[y[k]-1:ey[k],x[k]-1:ex[k],:] # tmp(dy(k):edy(k),dx(k):edx(k),:)=img(y(k):ey(k),x(k):ex(k),:);
121+
if tmp.shape[0]>0 and tmp.shape[1]>0 or tmp.shape[0]==0 and tmp.shape[1]==0: # if size(tmp,1)>0 && size(tmp,2)>0 || size(tmp,1)==0 && size(tmp,2)==0
122+
tempimg[:,:,:,k] = imResample2(tmp, (48, 48), 'bilinear') # tempimg(:,:,:,k)=imResample(tmp,[48 48],'bilinear');
123+
else:
124+
return np.empty() # total_boxes = []; return;
125+
tempimg = (tempimg-127.5)*0.0078125 # tempimg=(tempimg-127.5)*0.0078125;
126+
tempimg1 = np.transpose(tempimg, (3,1,0,2))
127+
out = onet(tempimg1) # out=ONet.forward({tempimg});
128+
out0 = np.transpose(out[0])
129+
out1 = np.transpose(out[1])
130+
out2 = np.transpose(out[2])
131+
score = out2[1,:] # score=squeeze(out{3}(2,:));
132+
points = out1 # points=out{2};
133+
ipass = np.where(score>threshold[2]) # pass=find(score>threshold(3));
134+
points = points[:,ipass[0]] # points=points(:,pass);
135+
total_boxes = np.hstack([total_boxes[ipass[0],0:4].copy(), np.expand_dims(score[ipass].copy(),1)]) # total_boxes=[total_boxes(pass,1:4) score(pass)'];
136+
mv = out0[:,ipass[0]] # mv=out{1}(:,pass);
137+
138+
w = total_boxes[:,2]-total_boxes[:,0]+1 # w=total_boxes(:,3)-total_boxes(:,1)+1;
139+
h = total_boxes[:,3]-total_boxes[:,1]+1 # h=total_boxes(:,4)-total_boxes(:,2)+1;
140+
points[0:5,:] = np.tile(w,(5, 1))*points[0:5,:] + np.tile(total_boxes[:,0],(5, 1))-1 # points(1:5,:)=repmat(w',[5 1]).*points(1:5,:)+repmat(total_boxes(:,1)',[5 1])-1;
141+
points[5:10,:] = np.tile(h,(5, 1))*points[5:10,:] + np.tile(total_boxes[:,1],(5, 1))-1 # points(6:10,:)=repmat(h',[5 1]).*points(6:10,:)+repmat(total_boxes(:,2)',[5 1])-1;
142+
if total_boxes.shape[0]>0: # if size(total_boxes,1)>0
143+
total_boxes = bbreg(total_boxes.copy(), np.transpose(mv)) # total_boxes=bbreg(total_boxes,mv(:,:)');
144+
pick = nms(total_boxes.copy(), 0.7, 'Min') # pick=nms(total_boxes,0.7,'Min');
145+
total_boxes = total_boxes[pick,:] # total_boxes=total_boxes(pick,:);
146+
points = points[:,pick] # points=points(:,pick);
147+
148+
return total_boxes, points
149+
150+
151+
# function [boundingbox] = bbreg(boundingbox,reg)
152+
def bbreg(boundingbox,reg):
153+
# calibrate bouding boxes
154+
if reg.shape[1]==1: # if size(reg,2)==1
155+
reg = np.reshape(reg, (reg.shape[2], reg.shape[3])) # reg=reshape(reg,[size(reg,3) size(reg,4)])';
156+
157+
w = boundingbox[:,2]-boundingbox[:,0]+1 # w=[boundingbox(:,3)-boundingbox(:,1)]+1;
158+
h = boundingbox[:,3]-boundingbox[:,1]+1 # h=[boundingbox(:,4)-boundingbox(:,2)]+1;
159+
b1 = boundingbox[:,0]+reg[:,0]*w
160+
b2 = boundingbox[:,1]+reg[:,1]*h
161+
b3 = boundingbox[:,2]+reg[:,2]*w
162+
b4 = boundingbox[:,3]+reg[:,3]*h
163+
boundingbox[:,0:4] = np.transpose(np.vstack([b1, b2, b3, b4 ])) # boundingbox(:,1:4)=[boundingbox(:,1)+reg(:,1).*w boundingbox(:,2)+reg(:,2).*h boundingbox(:,3)+reg(:,3).*w boundingbox(:,4)+reg(:,4).*h];
164+
return boundingbox
165+
166+
def generateBoundingBox(map, reg, scale, t):
167+
# use heatmap to generate bounding boxes
168+
stride=2
169+
cellsize=12
170+
171+
map = np.transpose(map) # map=map';
172+
dx1 = np.transpose(reg[:,:,0]) # dx1=reg(:,:,1)';
173+
dy1 = np.transpose(reg[:,:,1]) # dy1=reg(:,:,2)';
174+
dx2 = np.transpose(reg[:,:,2]) # dx2=reg(:,:,3)';
175+
dy2 = np.transpose(reg[:,:,3]) # dy2=reg(:,:,4)';
176+
y, x = np.where(map > t) # [y x]=find(map>=t);
177+
if y.shape[0]==1: # if size(y,1)==1
178+
# *** not checked ***
179+
# y=y';x=x';score=map(a)';dx1=dx1';dy1=dy1';dx2=dx2';dy2=dy2';
180+
y = np.transpose(y)
181+
x = np.transpose(x)
182+
dx1 = np.transpose(dx1)
183+
dy1 = np.transpose(dy1)
184+
dx2 = np.transpose(dx2)
185+
dy2 = np.transpose(dy2)
186+
score = map[(y,x)]
187+
else:
188+
score = map[(y,x)]
189+
190+
reg = np.transpose(np.vstack([ dx1[(y,x)], dy1[(y,x)], dx2[(y,x)], dy2[(y,x)] ])) # reg=[dx1(a) dy1(a) dx2(a) dy2(a)];
191+
if reg.size==0: # if isempty(reg)
192+
# *** not checked ***
193+
reg = np.empty((0,3)) # reg=reshape([],[0 3]);
194+
bb = np.transpose(np.vstack([y,x])) # boundingbox=[y x];
195+
q1 = np.fix((stride*bb+1)/scale)
196+
q2 = np.fix((stride*bb+cellsize-1+1)/scale)
197+
boundingbox = np.hstack([q1, q2, np.expand_dims(score,1), reg]) # boundingbox=[fix((stride*(boundingbox-1)+1)/scale) fix((stride*(boundingbox-1)+cellsize-1+1)/scale) score reg];
198+
return boundingbox, reg
199+
200+
# function pick = nms(boxes,threshold,type)
201+
def nms(boxes,threshold,type):
202+
if boxes.size==0: # if isempty(boxes)
203+
return np.empty((0,3)) ###### check dimension #######
204+
x1 = boxes[:,0]
205+
y1 = boxes[:,1]
206+
x2 = boxes[:,2]
207+
y2 = boxes[:,3]
208+
s = boxes[:,4]
209+
area = (x2-x1+1) * (y2-y1+1)
210+
I = np.argsort(s) # [vals, I] = sort(s);
211+
#vals = s[I]
212+
pick = np.zeros_like(s, dtype=np.int16) # s*0;
213+
counter = 0
214+
while I.size>0: # ~isempty(I)
215+
#last = I.size
216+
i = I[-1]
217+
pick[counter] = i
218+
counter += 1
219+
idx = I[0:-1]
220+
xx1 = np.maximum(x1[i], x1[idx]) # xx1 = max(x1(i), x1(I(1:last-1)));
221+
yy1 = np.maximum(y1[i], y1[idx]) # yy1 = max(y1(i), y1(I(1:last-1)));
222+
xx2 = np.minimum(x2[i], x2[idx]) # xx2 = min(x2(i), x2(I(1:last-1)));
223+
yy2 = np.minimum(y2[i], y2[idx]) # yy2 = min(y2(i), y2(I(1:last-1)));
224+
w = np.maximum(0.0, xx2-xx1+1) # w = max(0.0, xx2-xx1+1);
225+
h = np.maximum(0.0, yy2-yy1+1) # h = max(0.0, yy2-yy1+1);
226+
inter = w * h # inter = w.*h;
227+
if type is 'Min': # if strcmp(type,'Min')
228+
o = inter / np.minimum(area[i], area[idx]) # o = inter ./ min(area(i),area(I(1:last-1)));
229+
else:
230+
o = inter / (area[i] + area[idx] - inter) # o = inter ./ (area(i) + area(I(1:last-1)) - inter);
231+
I = I[np.where(o<=threshold)] # I = I(find(o<=threshold));
232+
pick = pick[0:counter] # pick = pick(1:(counter-1));
233+
return pick
234+
235+
# function [dy edy dx edx y ey x ex tmpw tmph] = pad(total_boxes,w,h)
236+
def pad(total_boxes, w, h):
237+
# compute the padding coordinates (pad the bounding boxes to square)
238+
tmpw = total_boxes[:,2]-total_boxes[:,0]+1 # tmpw=total_boxes(:,3)-total_boxes(:,1)+1;
239+
tmph = total_boxes[:,3]-total_boxes[:,1]+1 # tmph=total_boxes(:,4)-total_boxes(:,2)+1;
240+
numbox=total_boxes.shape[0]
241+
242+
dx = np.ones((numbox,1)) # dx=ones(numbox,1);
243+
dy = np.ones((numbox,1)) # dy=ones(numbox,1);
244+
edx=tmpw.copy()
245+
edy=tmph.copy()
246+
247+
x = total_boxes[:,0].copy() # x=total_boxes(:,1);
248+
y = total_boxes[:,1].copy() # y=total_boxes(:,2);
249+
ex = total_boxes[:,2].copy() # ex=total_boxes(:,3);
250+
ey = total_boxes[:,3].copy() # ey=total_boxes(:,4);
251+
252+
tmp = np.where(ex>w) # tmp=find(ex>w);
253+
edx[tmp] = np.expand_dims(-ex[tmp]+w+tmpw[tmp],1) # edx(tmp)=-ex(tmp)+w+tmpw(tmp);ex(tmp)=w;
254+
ex[tmp] = w
255+
256+
tmp = np.where(ey>h) # tmp=find(ey>h);
257+
edy[tmp] = np.expand_dims(-ey[tmp]+h+tmph[tmp],1) # edy(tmp)=-ey(tmp)+h+tmph(tmp);ey(tmp)=h;
258+
ey[tmp] = h
259+
260+
tmp = np.where(x<1) # tmp=find(x<1);
261+
dx[tmp] = np.expand_dims(2-x[tmp],1) # dx(tmp)=2-x(tmp);
262+
x[tmp] = 1
263+
264+
tmp = np.where(y<1) # tmp=find(y<1);
265+
dy[tmp] = np.expand_dims(2-y[tmp],1) # dy(tmp)=2-y(tmp);y(tmp)=1;
266+
y[tmp] = 1
267+
268+
return dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph
269+
270+
# function [bboxA] = rerec(bboxA)
271+
def rerec(bboxA):
272+
# convert bboxA to square
273+
#bboxB = bboxA[:,0:4] # bboxB=bboxA(:,1:4);
274+
h = bboxA[:,3]-bboxA[:,1] # h=bboxA(:,4)-bboxA(:,2);
275+
w = bboxA[:,2]-bboxA[:,0] # w=bboxA(:,3)-bboxA(:,1);
276+
l = np.maximum(w, h) # l=max([w h]')';
277+
bboxA[:,0] = bboxA[:,0]+w*0.5-l*0.5 # bboxA(:,1)=bboxA(:,1)+w.*0.5-l.*0.5;
278+
bboxA[:,1] = bboxA[:,1]+h*0.5-l*0.5 # bboxA(:,2)=bboxA(:,2)+h.*0.5-l.*0.5;
279+
bboxA[:,2:4] = bboxA[:,0:2] + np.transpose(np.tile(l,(2,1))) # bboxA(:,3:4)=bboxA(:,1:2)+repmat(l,[1 2]);
280+
return bboxA
281+
282+
def imResample2(img, sz, method):
283+
h=img.shape[0]
284+
w=img.shape[1]
285+
hs, ws = sz
286+
dx = float(w) / ws
287+
dy = float(h) / hs
288+
im_data = np.zeros((hs,ws,3))
289+
for a1 in range(0,hs):
290+
for a2 in range(0,ws):
291+
for a3 in range(0,3):
292+
im_data[a1,a2,a3] = img[floor(a1*dy),floor(a2*dx),a3]
293+
return im_data

Diff for: ‎src/align/detect_face.pyc

9.23 KB
Binary file not shown.

0 commit comments

Comments
 (0)