modules/softcascade/misc/scale_caltech.py

   1 #!/usr/bin/env python
   2
   3 import sys, os, os.path, glob, math, cv2, sft
   4 from datetime import datetime
   5 from optparse import OptionParser
   6 import re
   7 import numpy as np
   8
   9 def extractPositive(f, path, opath, octave, min_possible):
  10     newobj = re.compile("^lbl=\'(\w+)\'\s+str=(\d+)\s+end=(\d+)\s+hide=0$")
  11     pos    = re.compile("^pos\s=(\[[((\d+\.+\d*)|\s+|\;)]*\])$")
  12     occl   = re.compile("^occl\s*=(\[[0-1|\s]*\])$")
  13
  14     whole_mod_w = int(64  * octave) + 2 * int(20 * octave)
  15     whole_mod_h = int(128 * octave) + 2 * int(20 * octave)
  16
  17     goNext = 0
  18     start  = 0
  19     end    = 0
  20
  21     person_id = -1;
  22
  23     boxes = []
  24     occls = []
  25
  26     for l in f:
  27         m = newobj.match(l)
  28         if m is not None:
  29             if m.group(1) == "person":
  30                 goNext = 1
  31                 start = int(m.group(2))
  32                 end   = int(m.group(3))
  33                 person_id = person_id + 1
  34                 print m.group(1), person_id, start, end
  35             else:
  36                 goNext = 0
  37         else:
  38             m = pos.match(l)
  39             if m is not None:
  40                 if not goNext:
  41                     continue
  42                 strarr = re.sub(r"\s", ", ", re.sub(r"\;\s+(?=\])", "]", re.sub(r"\;\s+(?!\])", "],[", re.sub(r"(\[)(\d)", "\\1[\\2", m.group(1)))))
  43                 boxes = eval(strarr)
  44             else:
  45                 m = occl.match(l)
  46                 if m is not None:
  47                     occls = eval(re.sub(r"\s+(?!\])", ",", m.group(1)))
  48
  49                     if len(boxes) > 0 and len(boxes) == len(occls):
  50                         for idx, box in enumerate(boxes):
  51                             if occls[idx] == 1:
  52                                 continue
  53
  54                             x = box[0]
  55                             y = box[1]
  56                             w = box[2]
  57                             h = box[3]
  58
  59                             id = int(start) - 1 + idx
  60                             file = os.path.join(path, "I0%04d.jpg" % id)
  61
  62                             if (start + id) >= end or w < 10 or h < min_possible:
  63                                 continue
  64
  65                             mat = cv2.imread(file)
  66                             mat_h, mat_w, _ = mat.shape
  67
  68                             # let default height of person be 96.
  69                             scale = h / float(96)
  70                             rel_scale = scale / octave
  71
  72                             d_w = whole_mod_w * rel_scale
  73                             d_h = whole_mod_h * rel_scale
  74
  75                             tb = (d_h - h) / 2.0
  76                             lr = (d_w - w) / 2.0
  77
  78                             x = int(round(x - lr))
  79                             y = int(round(y - tb))
  80
  81                             w = int(round(w + lr * 2.0))
  82                             h = int(round(h + tb * 2.0))
  83
  84                             inner = [max(5, x), max(5, y), min(mat_w - 5, x + w), min(mat_h - 5, y + h) ]
  85                             cropped = mat[inner[1]:inner[3], inner[0]:inner[2], :]
  86
  87                             top     = int(max(0, 0 - y))
  88                             bottom  = int(max(0, y + h - mat_h))
  89                             left    = int(max(0, 0 - x))
  90                             right   = int(max(0, x + w - mat_w))
  91
  92                             if top < -d_h / 4.0 or bottom > d_h / 4.0 or left < -d_w / 4.0 or right > d_w / 4.0:
  93                                 continue
  94
  95                             cropped = cv2.copyMakeBorder(cropped, top, bottom, left, right, cv2.BORDER_REPLICATE)
  96                             resized = sft.resize_sample(cropped, whole_mod_w, whole_mod_h)
  97                             flipped = cv2.flip(resized, 1)
  98
  99                             cv2.imshow("resized", resized)
 100
 101                             c = cv2.waitKey(20)
 102                             if c == 27:
 103                                 exit(0)
 104
 105                             fname = re.sub(r"^.*\/(set[0-1]\d)\/(V0\d\d)\.(seq)/(I\d+).jpg$", "\\1_\\2_\\4", file)
 106                             fname = os.path.join(opath, fname + "_%04d." % person_id + "png")
 107                             fname_fl = os.path.join(opath, fname + "_mirror_%04d." % person_id + "png")
 108                             try:
 109                                 cv2.imwrite(fname, resized)
 110                                 cv2.imwrite(fname_fl, flipped)
 111                             except:
 112                                 print "something wrong... go next."
 113                                 pass
 114
 115 if __name__ == "__main__":
 116     parser = OptionParser()
 117     parser.add_option("-i", "--input", dest="input", metavar="DIRECTORY", type="string",
 118                        help="Path to the Caltech dataset folder.")
 119
 120     parser.add_option("-d", "--output-dir", dest="output", metavar="DIRECTORY", type="string",
 121                        help="Path to store data", default=".")
 122
 123     parser.add_option("-o", "--octave", dest="octave", type="float",
 124                        help="Octave for a dataset to be scaled", default="0.5")
 125
 126     parser.add_option("-m", "--min-possible", dest="min_possible", type="int",
 127                        help="Minimum possible height for positive.", default="64")
 128
 129     (options, args) = parser.parse_args()
 130
 131     if not options.input:
 132         parser.error("Caltech dataset folder is required.")
 133
 134     opath = os.path.join(options.output, datetime.now().strftime("raw_ge64_cr_mirr_ts" + "-%Y-%m-%d-%H-%M-%S"))
 135     os.mkdir(opath)
 136
 137     gl = glob.iglob( os.path.join(options.input, "set[0][0]/V0[0-9][0-9].txt"))
 138     for each in gl:
 139         path, ext = os.path.splitext(each)
 140         path = path + ".seq"
 141         print path
 142         extractPositive(open(each), path, opath, options.octave, options.min_possible)