def stitch_rects(all_rects, tau=0.25):
"""
Implements the stitching procedure discussed in the paper.
Complicated, but we find that it does better than simpler versions
and generalizes well across widely varying box sizes.
Input:
all_rects : 2d grid with each cell containing a vector of PyRects
"""
for row in all_rects:
assert len(row) == len(all_rects[0])
cdef vector[vector[vector[Rect]]] c_rects
cdef vector[vector[Rect]] c_row
cdef vector[Rect] c_column
for i, row in enumerate(all_rects):
c_rects.push_back(c_row)
for j, column in enumerate(row):
c_rects[i].push_back(c_column)
for py_rect in column:
c_rects[i][j].push_back(
Rect(
py_rect.cx,
py_rect.cy,
py_rect.width,
py_rect.height,
py_rect.confidence)
)
cdef vector[Rect] acc_rects;
thresholds = [(.80, 1.0),
(.70, 0.9),
(.60, 0.8),
(.50, 0.7),
(.40, 0.6),
(.30, 0.5),
(.20, 0.4),
(.10, 0.3),
(.05, 0.2),
(.02, 0.1),
(.005, 0.04),
(.001, 0.01),
]
t_conf_alphas = [(tau, 1.0),
#(1 - (1 - tau) * 0.75, 0.5),
#(1 - (1 - tau) * 0.5, 0.1),
#(1 - (1 - tau) * 0.25, 0.005),
]
for t, conf_alpha in t_conf_alphas:
for lower_t, upper_t in thresholds:
if lower_t * conf_alpha > 0.0001:
filter_rects(c_rects, &acc_rects, lower_t * conf_alpha,
upper_t * conf_alpha, t, conf_alpha)
py_acc_rects = []
for i in range(acc_rects.size()):
acc_rect = PyRect(
acc_rects[i].cx_,
acc_rects[i].cy_,
acc_rects[i].width_,
acc_rects[i].height_,
acc_rects[i].confidence_)
acc_rect.true_confidence = acc_rects[i].true_confidence_
py_acc_rects.append(acc_rect)
return py_acc_rects
