Preprocessing¶
src/label.py¶
Class: Label¶
import numpy as np
from os.path import isfile
class Label:
def __init__(self,cl=-1,tl=np.array([0.,0.]),br=np.array([0.,0.]),prob=None):
self.__tl = tl
self.__br = br
self.__cl = cl
self.__prob = prob
def __str__(self):
return 'Class: %d, top_left(x:%f,y:%f), bottom_right(x:%f,y:%f)' % (self.__cl, self.__tl[0], self.__tl[1], self.__br[0], self.__br[1])
def copy(self):
return Label(self.__cl,self.__tl,self.__br)
def wh(self): return self.__br-self.__tl
def cc(self): return self.__tl + self.wh()/2
def tl(self): return self.__tl
def br(self): return self.__br
def tr(self): return np.array([self.__br[0],self.__tl[1]])
def bl(self): return np.array([self.__tl[0],self.__br[1]])
def cl(self): return self.__cl
def area(self): return np.prod(self.wh())
def prob(self): return self.__prob
def set_class(self,cl):
self.__cl = cl
def set_tl(self,tl):
self.__tl = tl
def set_br(self,br):
self.__br = br
def set_wh(self,wh):
cc = self.cc()
self.__tl = cc - .5*wh
self.__br = cc + .5*wh
def set_prob(self,prob):
self.__prob = prob
def lread(file_path,label_type=Label):
if not isfile(file_path):
return []
objs = []
with open(file_path,'r') as fd:
for line in fd:
v = line.strip().split()
cl = int(v[0])
ccx,ccy = float(v[1]),float(v[2])
w,h = float(v[3]),float(v[4])
prob = float(v[5]) if len(v) == 6 else None
cc = np.array([ccx,ccy])
wh = np.array([w,h])
objs.append(label_type(cl,cc-wh/2,cc+wh/2,prob=prob))
return objs
def lwrite(file_path,labels,write_probs=True):
with open(file_path,'w') as fd:
for l in labels:
cc,wh,cl,prob = (l.cc(),l.wh(),l.cl(),l.prob())
if prob != None and write_probs:
fd.write('%d %f %f %f %f %f\n' % (cl,cc[0],cc[1],wh[0],wh[1],prob))
else:
fd.write('%d %f %f %f %f\n' % (cl,cc[0],cc[1],wh[0],wh[1]))
def dknet_label_conversion(R,img_width,img_height):
WH = np.array([img_width,img_height],dtype=float)
L = []
for r in R:
center = np.array(r[2][:2])/WH
wh2 = (np.array(r[2][2:])/WH)*.5
L.append(Label(ord(r[0]),tl=center-wh2,br=center+wh2,prob=r[1]))
return L
Class: Shape¶
class Shape():
def __init__(self,pts=np.zeros((2,0)),max_sides=4,text=''):
self.pts = pts
self.max_sides = max_sides
self.text = text
def isValid(self):
return self.pts.shape[1] > 2
def write(self,fp):
fp.write('%d,' % self.pts.shape[1])
ptsarray = self.pts.flatten()
fp.write(''.join([('%f,' % value) for value in ptsarray]))
fp.write('%s,' % self.text)
fp.write('\n')
def read(self,line):
data = line.strip().split(',')
ss = int(data[0])
values = data[1:(ss*2 + 1)]
text = data[(ss*2 + 1)] if len(data) >= (ss*2 + 2) else ''
self.pts = np.array([float(value) for value in values]).reshape((2,ss))
self.text = text
def readShapes(path, obj_type=Shape):
shapes = []
with open(path) as fp:
for line in fp:
shape = obj_type()
shape.read(line)
shapes.append(shape)
return shapes
def writeShapes(path, shapes):
if len(shapes):
with open(path, 'w') as fp:
for shape in shapes:
if shape.isValid():
shape.write(fp)
Link: train-detector.py
src/sampler.py¶
Import libraries¶
import cv2
import numpy as np
import random
from src.utils import im2single, getWH, hsv_transform, IOU_centre_and_dims
from src.label import Label
from src.projection_utils import perspective_transform, find_T_matrix, getRectPts
Data augmentation¶
def pts2ptsh(pts):
return np.matrix(np.concatenate((pts, np.ones((1, pts.shape[1]))), 0))
def project(I, T, pts, dim):
ptsh = np.matrix(np.concatenate((pts, np.ones((1, 4))), 0))
ptsh = np.matmul(T, ptsh)
ptsh = ptsh / ptsh[2]
ptsret = ptsh[:2]
ptsret = ptsret / dim
Iroi = cv2.warpPerspective(I, T, (dim, dim), borderValue=.0, flags=cv2.INTER_LINEAR)
return Iroi, ptsret
def flip_image_and_pts(I, pts):
"""Flip an image
Args:
I:
pts:
Returns:
I:
pts:
"""
I = cv2.flip(I, 1)
pts[0] = 1. - pts[0]
idx = [1, 0, 3, 2]
pts = pts[..., idx]
return I, pts
def augment_sample(I, pts, dim):
"""Data augmentation
Args:
I:
pts:
dim:
Returns:
Iroi:
llp:
pts:
"""
maxsum, maxangle = 120, np.array([80., 80., 45.])
angles = np.random.rand(3) * maxangle
if angles.sum() > maxsum:
angles = (angles / angles.sum()) * (maxangle/maxangle.sum())
I = im2single(I)
iwh = getWH(I.shape)
# Aspect-ratio: the LP aspect-ratio is randomly set in the interval [2, 4]
# to accommodate sizes from different regions
whratio = random.uniform(2., 4.)
wsiz = random.uniform(dim * .2, dim * 1.)
hsiz = wsiz / whratio
dx = random.uniform(0., dim - wsiz)
dy = random.uniform(0., dim - hsiz)
# Rectification: the entire image is rectified based on the LP annotation,
# assuming that the LP lies on a plane
pph = getRectPts(dx, dy, dx+wsiz, dy+hsiz)
pts = pts * iwh.reshape((2, 1))
T = find_T_matrix(pts2ptsh(pts), pph)
# Rotation: a 3D rotation with randomly chosen angles is performed,
# to account for a wide range of camera setups
H = perspective_transform((dim, dim), angles=angles)
H = np.matmul(H, T)
Iroi, pts = project(I, H, pts, dim)
# Colorspace: slight modifications in the HSV colorspace
hsv_mod = np.random.rand(3).astype('float32')
hsv_mod = (hsv_mod - .5)*.3
hsv_mod[0] *= 360
Iroi = hsv_transform(Iroi, hsv_mod)
Iroi = np.clip(Iroi, 0., 1.)
pts = np.array(pts)
# Mirroring 50% chance
if random.random() > .5:
Iroi, pts = flip_image_and_pts(Iroi, pts)
tl, br = pts.min(1), pts.max(1)
llp = Label(0, tl, br)
return Iroi, llp, pts
Link: im2single, process_data_item
Label ￫ Output vector¶
def labels2output_map(label, lppts, dim, stride):
side = ((float(dim) + 40.) / 2.) / stride # 7.75 when dim = 208 and stride = 16
outsize = dim / stride
Y = np.zeros((outsize, outsize, 2 * 4 + 1), dtype='float32')
MN = np.array([outsize, outsize])
WH = np.array([dim, dim], dtype=float)
tlx, tly = np.floor(np.maximum(label.tl(), 0.) * MN).astype(int).tolist()
brx, bry = np.ceil (np.minimum(label.br(), 1.) * MN).astype(int).tolist()
for x in range(tlx, brx):
for y in range(tly, bry):
mn = np.array([float(x) + .5, float(y) + .5])
iou = IOU_centre_and_dims(mn / MN, label.wh(), label.cc(), label.wh())
if iou > .5:
p_WH = lppts * WH.reshape((2, 1))
p_MN = p_WH / stride
p_MN_center_mn = p_MN - mn.reshape((2, 1))
p_side = p_MN_center_mn / side
Y[y, x, 0] = 1.
Y[y, x, 1:] = p_side.T.flatten()
return Y
Link: process_data_item