Open Source Computer Vision Library https://opencv.org/
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

281 lines
8.6 KiB

#!/usr/bin/env python
import cv2, re, glob
import numpy as np
import matplotlib.pyplot as plt
from itertools import izip
""" Convert numPy matrices with rectangles and confidences to sorted list of detections."""
def convert2detections(rects, confs, crop_factor = 0.125):
if rects is None:
return []
dts = zip(*[rects.tolist(), confs.tolist()])
dts = zip(dts[0][0], dts[0][1])
dts = [Detection(r,c) for r, c in dts]
dts.sort(lambda x, y : -1 if (x.conf - y.conf) > 0 else 1)
for dt in dts:
dt.crop(crop_factor)
return dts
""" Create new instance of soft cascade."""
def cascade(min_scale, max_scale, nscales, f):
# where we use nms cv::SoftCascadeDetector::DOLLAR == 2
c = cv2.softcascade_Detector(min_scale, max_scale, nscales, 2)
xml = cv2.FileStorage(f, 0)
dom = xml.getFirstTopLevelNode()
assert c.load(dom)
return c
""" Compute prefix sum for en array."""
def cumsum(n):
cum = []
y = 0
for i in n:
y += i
cum.append(y)
return cum
""" Compute x and y arrays for ROC plot."""
def computeROC(confidenses, tp, nannotated, nframes, ignored):
confidenses, tp, ignored = zip(*sorted(zip(confidenses, tp, ignored), reverse = True))
fp = [(1 - x) for x in tp]
fp = [(x - y) for x, y in izip(fp, ignored)]
fp = cumsum(fp)
tp = cumsum(tp)
miss_rate = [(1 - x / (nannotated + 0.000001)) for x in tp]
fppi = [x / float(nframes) for x in fp]
return fppi, miss_rate
""" Crop rectangle by factor."""
def crop_rect(rect, factor):
val_x = factor * float(rect[2])
val_y = factor * float(rect[3])
x = [int(rect[0] + val_x), int(rect[1] + val_y), int(rect[2] - 2.0 * val_x), int(rect[3] - 2.0 * val_y)]
return x
""" Initialize plot axises."""
def initPlot(name):
plt.xlabel("fppi")
plt.ylabel("miss rate")
plt.title(name)
plt.grid(True)
plt.xscale('log')
plt.yscale('log')
""" Draw plot."""
def plotLogLog(fppi, miss_rate, c):
plt.loglog(fppi, miss_rate, color = c, linewidth = 2)
""" Show resulted plot."""
def showPlot(file_name, labels):
plt.axis((pow(10, -3), pow(10, 1), .035, 1))
plt.yticks( [0.05, 0.1, 0.2, 0.3, 0.4, 0.5, 0.64, 0.8, 1], ['.05', '.10', '.20', '.30', '.40', '.50', '.64', '.80', '1'] )
plt.legend(labels, loc = "lower left")
plt.savefig(file_name)
plt.show()
""" Filter true positives and ignored detections for cascade detector output."""
def match(gts, dts):
matches_gt = [0]*len(gts)
matches_dt = [0]*len(dts)
matches_ignore = [0]*len(dts)
if len(gts) == 0:
return matches_dt, matches_ignore
# Cartesian product for each detection BB_dt with each BB_gt
overlaps = [[dt.overlap(gt) for gt in gts]for dt in dts]
for idx, row in enumerate(overlaps):
imax = row.index(max(row))
# try to match ground truth
if (matches_gt[imax] == 0 and row[imax] > 0.5):
matches_gt[imax] = 1
matches_dt[idx] = 1
for idx, dt in enumerate(dts):
# try to math ignored
if matches_dt[idx] == 0:
row = gts
row = [i for i in row if (i[3] - i[1]) < 53 or (i[3] - i[1]) > 256]
for each in row:
if dts[idx].overlapIgnored(each) > 0.5:
matches_ignore[idx] = 1
return matches_dt, matches_ignore
""" Draw detections or ground truth on image."""
def draw_rects(img, rects, color, l = lambda x, y : x + y):
if rects is not None:
for x1, y1, x2, y2 in rects:
cv2.rectangle(img, (x1, y1), (l(x1, x2), l(y1, y2)), color, 2)
def draw_dt(img, dts, color, l = lambda x, y : x + y):
if dts is not None:
for dt in dts:
bb = dt.bb
x1, y1, x2, y2 = dt.bb[0], dt.bb[1], dt.bb[2], dt.bb[3]
cv2.rectangle(img, (x1, y1), (l(x1, x2), l(y1, y2)), color, 2)
class Detection:
def __init__(self, bb, conf):
self.bb = bb
self.conf = conf
self.matched = False
def crop(self, factor):
self.bb = crop_rect(self.bb, factor)
# we use rect-style for dt and box style for gt. ToDo: fix it
def overlap(self, b):
a = self.bb
w = min( a[0] + a[2], b[2]) - max(a[0], b[0]);
h = min( a[1] + a[3], b[3]) - max(a[1], b[1]);
cross_area = 0.0 if (w < 0 or h < 0) else float(w * h)
union_area = (a[2] * a[3]) + ((b[2] - b[0]) * (b[3] - b[1])) - cross_area;
return cross_area / union_area
# we use rect-style for dt and box style for gt. ToDo: fix it
def overlapIgnored(self, b):
a = self.bb
w = min( a[0] + a[2], b[2]) - max(a[0], b[0]);
h = min( a[1] + a[3], b[3]) - max(a[1], b[1]);
cross_area = 0.0 if (w < 0 or h < 0) else float(w * h)
self_area = (a[2] * a[3]);
return cross_area / self_area
def mark_matched(self):
self.matched = True
"""Parse INPIA annotation format"""
def parse_inria(ipath, f):
bbs = []
path = None
for l in f:
box = None
if l.startswith("Bounding box"):
b = [x.strip() for x in l.split(":")[1].split("-")]
c = [x[1:-1].split(",") for x in b]
d = [int(x) for x in sum(c, [])]
bbs.append(d)
if l.startswith("Image filename"):
path = l.split('"')[-2]
return Sample(path, bbs)
def glob_set(pattern):
return [__n for __n in glob.iglob(pattern)]
""" Parse ETH idl file. """
def parse_idl(f):
map = {}
for l in open(f):
l = re.sub(r"^\"left\/", "{\"", l)
l = re.sub(r"\:", ":[", l)
l = re.sub(r"(\;|\.)$", "]}", l)
map.update(eval(l))
return map
""" Normalize detection box to unified aspect ration."""
def norm_box(box, ratio):
middle = float(box[0] + box[2]) / 2.0
new_half_width = float(box[3] - box[1]) * ratio / 2.0
return (int(round(middle - new_half_width)), box[1], int(round(middle + new_half_width)), box[3])
""" Process array of boxes."""
def norm_acpect_ratio(boxes, ratio):
return [ norm_box(box, ratio) for box in boxes]
""" Filter detections out of extended range. """
def filter_for_range(boxes, scale_range, ext_ratio):
boxes = norm_acpect_ratio(boxes, 0.5)
boxes = [b for b in boxes if (b[3] - b[1]) > scale_range[0] / ext_ratio]
boxes = [b for b in boxes if (b[3] - b[1]) < scale_range[1] * ext_ratio]
return boxes
""" Resize sample for training."""
def resize_sample(image, d_w, d_h):
h, w, _ = image.shape
if (d_h < h) or (d_w < w):
ratio = min(d_h / float(h), d_w / float(w))
kernel_size = int( 5 / (2 * ratio))
sigma = 0.5 / ratio
image_to_resize = cv2.filter2D(image, cv2.CV_8UC3, cv2.getGaussianKernel(kernel_size, sigma))
interpolation_type = cv2.INTER_AREA
else:
image_to_resize = image
interpolation_type = cv2.INTER_CUBIC
return cv2.resize(image_to_resize,(d_w, d_h), None, 0, 0, interpolation_type)
newobj = re.compile("^lbl=\'(\w+)\'\s+str=(\d+)\s+end=(\d+)\s+hide=0$")
class caltech:
@staticmethod
def extract_objects(f):
objects = []
tmp = []
for l in f:
if newobj.match(l) is not None:
objects.append(tmp)
tmp = []
tmp.append(l)
return objects[1:]
@staticmethod
def parse_header(f):
_ = f.readline() # skip first line (version string)
head = f.readline()
(nFrame, nSample) = re.search(r'nFrame=(\d+) n=(\d+)', head).groups()
return (int(nFrame), int(nSample))
@staticmethod
def parse_pos(l):
pos = re.match(r'^posv?\s*=(\[[\d\s\.\;]+\])$', l).group(1)
pos = re.sub(r"(\[)(\d)", "\\1[\\2", pos)
pos = re.sub(r"\s", ", ", re.sub(r"\;\s+(?=\])", "]", re.sub(r"\;\s+(?!\])", "],[", pos)))
return eval(pos)
@staticmethod
def parse_occl(l):
occl = re.match(r'^occl\s*=(\[[\d\s\.\;]+\])$', l).group(1)
occl = re.sub(r"\s(?!\])", ",", occl)
12 years ago
return eval(occl)
def parse_caltech(f):
(nFrame, nSample) = caltech.parse_header(f)
objects = caltech.extract_objects(f)
annotations = [[] for i in range(nFrame)]
for obj in objects:
(type, start, end) = re.search(r'^lbl=\'(\w+)\'\s+str=(\d+)\s+end=(\d+)\s+hide=0$', obj[0]).groups()
print type, start, end
start = int(start) -1
end = int(end)
pos = caltech.parse_pos(obj[1])
posv = caltech.parse_pos(obj[2])
occl = caltech.parse_occl(obj[3])
for idx, (p, pv, oc) in enumerate(zip(*[pos, posv, occl])):
annotations[start + idx].append((type, p, oc, pv))
return annotations