def get_bezier_coefficient(self):
Mtk = lambda n, t, k: t**k * (1 - t)**(n - k) * n_over_k(n, k)
BezierCoeff = lambda ts: [[
Mtk(self.num_point - 1, t, k) for k in range(self.num_point)
] for t in ts]
return BezierCoeff
def overlay_bezier(self, image, predictions):
"""
Adds the predicted bezier on top of the image
"""
labels = predictions.get_field("labels")
# print('predictions', predictions.fields())
beziers = predictions.get_field("beziers")
colors = self.compute_colors_for_labels(labels).tolist()
Mtk = lambda n, t, k: t**k * (1 - t)**(n - k) * n_over_k(n, k)
BezierCoeff = lambda ts: [[Mtk(3, t, k) for k in range(4)] for t in ts]
for bez_pts, color in zip(beziers, colors):
bez_pts = bez_pts.to(torch.int64)
assert (
len(bez_pts) == 16
), 'The numbr of bezier control points must be 8, but got {}'.format(
len(bez_pts))
s1_bezier = bez_pts[:8].reshape((4, 2))[:, [1, 0]]
s2_bezier = bez_pts[8:].reshape((4, 2))[:, [1, 0]]
tpbtp = tuple(map(tuple, np.int32(s1_bezier)))
tpbbm = tuple(map(tuple, np.int32(s2_bezier)))
t_plot = np.linspace(0, 1, 81)
Bezier_top = np.array(BezierCoeff(t_plot)).dot(s1_bezier)
Bezier_bottom = np.array(BezierCoeff(t_plot)).dot(s2_bezier)
# use np.int32 for cv2 draw func
image = cv2.polylines(image, np.int32([Bezier_top]), False,
(0, 0, 255), 3)
image = cv2.polylines(image, np.int32([Bezier_bottom]), False,
(0, 0, 255), 3)
image = cv2.line(image, tpbtp[0], tpbbm[0], (0, 0, 255), 3)
image = cv2.line(image, tpbtp[3], tpbbm[3], (0, 0, 255), 3)
return image
def bezier_to_poly(bez, rec, rec_type):
Mtk = lambda n, t, k: t**k * (1-t)**(n-k) * n_over_k(n,k)
BezierCoeff = lambda ts: [[Mtk(3,t,k) for k in range(4)] for t in ts]
poly = []
assert(len(bez) == 16), 'The numbr of bezier control points must be 8, but got {}'.format(int(len(bez_pts)/2))
s1_bezier = bez[:8].reshape((4,2))[:, [1, 0]]
s2_bezier = bez[8:].reshape((4,2))[:, [1, 0]]
tpbtp = tuple(map(tuple, np.int32(s1_bezier)))
tpbbm = tuple(map(tuple, np.int32(s2_bezier)))
t_plot = np.linspace(0, 1, 20)
Bezier_top = np.array(BezierCoeff(t_plot)).dot(s1_bezier)
Bezier_bottom = np.array(BezierCoeff(t_plot)).dot(s2_bezier)
poly = Bezier_top.tolist()
bottom = Bezier_bottom.tolist()
bottom.reverse()
poly.extend(bottom)
if rec_type == "ctc":
last_char = False
s = ''
for c in rec:
c = int(c)
if c < 95:
if last_char != c:
s += CTLABELS[c]
last_char = c
elif c == 95:
s += u'口'
else:
last_char = False
elif rec_type == "attention":
s = ''
for c in rec:
c = int(c)
if c < 95:
s += CTLABELS[c]
elif c == 95:
s += u'口'
return poly, s
@Time : 2020/8/18 下午2:10
@Author : yizuotian
@Description : 转换windows_label_tool工具的标注
windows_label_tool 格式eg:
"""
import math
import numpy as np
import torch
from scipy.special import comb as n_over_k
from shapely.geometry import *
from sklearn.linear_model import LinearRegression
from sklearn.metrics import mean_squared_error
from torch import nn
Mtk = lambda n, t, k: t ** k * (1 - t) ** (n - k) * n_over_k(n, k)
BezierCoeff = lambda ts: [[Mtk(3, t, k) for k in range(4)] for t in ts]
class Bezier(nn.Module):
def __init__(self, ps, ctps):
"""
ps: numpy array of points
"""
super(Bezier, self).__init__()
self.x1 = nn.Parameter(torch.as_tensor(ctps[0], dtype=torch.float64))
self.x2 = nn.Parameter(torch.as_tensor(ctps[2], dtype=torch.float64))
self.y1 = nn.Parameter(torch.as_tensor(ctps[1], dtype=torch.float64))
self.y2 = nn.Parameter(torch.as_tensor(ctps[3], dtype=torch.float64))
self.x0 = ps[0, 0]
self.x3 = ps[-1, 0]
