def get_features(im):
f = np.zeros((im.shape[0], im.shape[1], 28))
if len(im.shape) == 2:
f[:, :, 0] = im - 0.5
f[1:-1, 1:-1, 1:] = pyhog.features_pedro(color.gray2rgb(im),
1)[:, :, :27]
else:
f[:, :, 0] = color.rgb2gray(im_patch) - 0.5
f[1:-1, 1:-1, 1:] = pyhog.features_pedro(color.gray2rgb(im),
1)[:, :, :27]
return f
def get_feature_map(im_patch):
out = np.zeros((im_patch.shape[0], im_patch.shape[1], 28))
if len(im_patch.shape) == 2:
out[:, :, 0] = im_patch - 0.5
temp = pyhog.features_pedro(skic.gray2rgb(im_patch), 1)
out[1:-1, 1:-1, 1:] = temp[:, :, :27]
else:
out[:, :, 0] = skic.rgb2gray(im_patch) - 0.5
temp = pyhog.features_pedro(skic.gray2rgb(im_patch), 1)
out[1:-1, 1:-1, 1:] = temp[:, :, :27]
return out
def get_scale_sample(im, pos, base_target_size, scale_factors, cos_window,
scale_model_size):
n = len(scale_factors)
for i in range(n):
patch_size = np.floor(base_target_size * scale_factors[i])
xs = np.floor(pos[1]) + np.arange(patch_size[1]) - np.floor(
patch_size[1] / 2)
ys = np.floor(pos[0]) + np.arange(patch_size[0]) - np.floor(
patch_size[0] / 2)
x_start = np.maximum(0, np.floor(pos[1]) - np.floor(patch_size[1] / 2))
y_start = np.maximum(0, np.floor(pos[0]) - np.floor(patch_size[0] / 2))
x_stop = np.minimum(
im.shape[1] - 1,
np.floor(pos[1]) + patch_size[1] - np.floor(patch_size[1] / 2))
y_stop = np.minimum(
im.shape[0] - 1,
np.floor(pos[0]) + patch_size[0] - np.floor(patch_size[0] / 2))
im_patch = im[y_start.astype(np.int32):y_stop.astype(np.int32),
x_start.astype(np.int32):x_stop.astype(np.int32)]
im_patch_resized = transform.resize(im_patch, scale_model_size)
temp_hog = pyhog.features_pedro(
repeat_to_third_dim(im_patch_resized, 3), 4)
temp = temp_hog[:, :, :32]
if i == 0:
features = np.zeros((temp.size, n))
features[:, i] = temp.ravel() * cos_window[i]
return features
def get_scale_sample(im, pos, base_target_sz, scaleFactors, scale_window,
scale_model_sz):
nScales = len(scaleFactors)
for s in range(nScales):
patch_sz = np.floor(base_target_sz * scaleFactors[s])
xs = np.floor(pos[1]) + np.arange(patch_sz[1]) - \
np.floor(patch_sz[1] / 2)
ys = np.floor(pos[0]) + np.arange(patch_sz[0]) - \
np.floor(patch_sz[0] / 2)
# # check for out-of-bounds coordinates, and set them to the values at
# # the borders
# xs[xs < 0] = 0
# ys[ys < 0] = 0
# xs[xs > im.shape[1] - 1] = im.shape[1] - 1
# ys[ys > im.shape[0] - 1] = im.shape[0] - 1
#
# # extract image
# im_patch = im[ys, xs, :]
x_start = np.maximum(0, np.floor(pos[1]) - np.floor(patch_sz[1] / 2))
y_start = np.maximum(0, np.floor(pos[0]) - np.floor(patch_sz[0] / 2))
x_stop = np.minimum(
im.shape[1] - 1,
np.floor(pos[1]) + patch_sz[1] - np.floor(patch_sz[1] / 2))
y_stop = np.minimum(
im.shape[0] - 1,
np.floor(pos[0]) + patch_sz[0] - np.floor(patch_sz[0] / 2))
# extract image
im_patch = im[x_start.astype(np.int32):x_stop.astype(np.int32),
y_start.astype(np.int32):y_stop.astype(np.int32)]
# resize image to model size
im_patch_resized = skit.resize(im_patch, scale_model_sz)
temp_hog = pyhog.features_pedro(
repeat_to_third_dim(im_patch_resized, 3), 4)
temp = temp_hog[:, :, :32]
if s == 0:
out = np.zeros((temp.size, nScales))
# window
out[:, s] = temp.ravel() * scale_window[s]
return out
def get_scale_sample(im, pos, base_target_sz, scaleFactors,
scale_window, scale_model_sz):
nScales = len(scaleFactors)
for s in range(nScales):
patch_sz = np.floor(base_target_sz * scaleFactors[s])
xs = np.floor(pos[1]) + np.arange(patch_sz[1]) - \
np.floor(patch_sz[1] / 2)
ys = np.floor(pos[0]) + np.arange(patch_sz[0]) - \
np.floor(patch_sz[0] / 2)
x_start = np.maximum(0, np.floor(pos[1]) - np.floor(patch_sz[1] / 2))
y_start = np.maximum(0, np.floor(pos[0]) - np.floor(patch_sz[0] / 2))
x_stop = np.minimum(im.shape[1] - 1,
np.floor(pos[1]) +
patch_sz[1] -
np.floor(patch_sz[1] / 2))
y_stop = np.minimum(im.shape[0] - 1,
np.floor(pos[0]) +
patch_sz[0] -
np.floor(patch_sz[0] / 2))
# extract image
im_patch = im[y_start.astype(np.int32):y_stop.astype(np.int32),
x_start.astype(np.int32):x_stop.astype(np.int32)]
# resize image to model size
im_patch_resized = skit.resize(im_patch, scale_model_sz)
temp_hog = pyhog.features_pedro(
repeat_to_third_dim(im_patch_resized, 3), 4)
temp = temp_hog[:, :, :32]
if s == 0:
out = np.zeros((temp.size, nScales))
# window
out[:, s] = temp.ravel() * scale_window[s]
return out
if not args.hog:
# normalize the image
print "Average grayvalue: %f" % (np.mean(imgraw,axis=None))
img = np.zeros(imgraw.shape)
if img.ndim==3:
img[:,:,0] = (imgraw[:,:,0] - np.mean(imgraw[:,:,0],axis=None)) / 255.0
img[:,:,1] = (imgraw[:,:,1] - np.mean(imgraw[:,:,1],axis=None)) / 255.0
img[:,:,2] = (imgraw[:,:,2] - np.mean(imgraw[:,:,2],axis=None)) / 255.0
else:
img = (imgraw - np.mean(imgraw,axis=None)) / 255.0
else:
# use HOG image
img = pyhog.features_pedro(imgraw, 8)
print "Dimensions of the image: ", img.shape
m = re.match('^(\d+):(\d+):(\d+)$', args.ps)
if m:
p = np.arange ( int(m.group(1)), int(m.group(3)), int(m.group(2)) )
patchSizes = np.vstack ( (p, p) ).T
else:
m = re.match('^(\d+)x(\d+)$', args.ps)
if m:
patchSizes = np.zeros([1,2])
patchSizes[0,0] = int(m.group(1))
patchSizes[0,1] = int(m.group(2))
else:
print "Wrong format given for the patch size"
def feature_pyramid(img):
"""
Compute feature pyramid for different image scales
Returns:
-------
tuple
-list, each element is a np.array which represents HoG feature
-np.array, each element is double-precision number which represents scale
"""
#import scipy.ndimage.interpolation as npimage
#it seems that using npimage.zoom brings some artifitials,
#Unfortunately, scipy.misc.imresize only return uint8 array.
from scipy.misc import imresize #however, still use imresize,
#since it's 30x faster
from pyhog import features_pedro
#change data type to float64
#img = img.astype(np.float64, copy=False)/255.0
MAXLEVELS = 200
MINIDIMENSION = 5
interval = detect_levels_per_octave
sc = 2 ** (1 / interval)
feat = []
scale = np.zeros((MAXLEVELS), dtype=float)
#Start at detect_max_scale, and keep going down by the increment sc, until
#we reach MAXLEVELS or detect_min_scale
for i in range(MAXLEVELS):
scaler = detect_max_scale / (sc ** (i))
if scaler < detect_min_scale:
return (feat, scale)
scale[i] = scaler
scaled = imresize(img, scaler).astype(np.float64, copy=False)/255.0
#if minimum dimensions is less than or equal to 5, return
if min(scaled.shape[0], scaled.shape[1]) <= MINIDIMENSION:
scale = scale[np.where(scale > 0.0)]
return (feat, scale)
feature = features_pedro(scaled, sbin)
#if we get zero size feature, backtrack one, and dont produce any
#more levels
if min(feature.shape[0], feature.shape[1]) == 0:
scale = scale[:,:i]
return (feat, scale)
#recover lost bin
feature = np.lib.pad(feature,((1,1),(1,1),(0,0)),
'constant',constant_values=(0.0))
feat.append(feature)
#if max dimensions is less than or equal to 5, return
if max(feature.shape[0], feature.shape[1]) <= MINIDIMENSION:
scale = scale[np.where(scale > 0.0)]
return (feat, scale)
return (feat, scale)
# -*- coding: utf-8 -*-
"""
Created on Fri Apr 1 11:02:37 2016
@author: liangfu
"""
import pyhog
import numpy as np
import matplotlib.pyplot as pl
from scipy.misc import imread, imrotate
import features_pedro_py
img = (imread('../VOC2007/JPEGImages/000015.jpg').astype(np.float64)/255.0)
pl.imshow(img)
pl.axis('off')
hog = pyhog.features_pedro(img, 30)
hog.shape
def main():
import pdb
pdb.set_trace()
im = Image.open("testImage.jpg")
hogbin = pyhog.features_pedro(im, 20)
pyhog.hog_picture(hogbin)
