def extract_hsv_features(ifile, pfile):
""" Given img file, and patched img file, return hue-sat 2Dhist and
val hist for each patch (10x10 + 10 = 110 dim feats) """
img = get_RGB(ifile)
p_img = get_RGB(pfile)
hsv_img = cv2.cvtColor(img, cv2.COLOR_RGB2HSV)
pi = Image.open(pfile)
n_patches = defaultdict(int)
for pixel in pi.getdata():
n_patches[pixel] += 1
if n_patches is None:
print("\nNo patches:", pfile)
return None
hsv_feats = []
patch_colors = []
for p in n_patches:
patch_pixels = np.where((p_img == p).all(axis=2))
hsv_values = hsv_img[patch_pixels[0], patch_pixels[1], :]
hs, _, _ = np.histogram2d(hsv_values[:, 0], hsv_values[:, 1], [10, 10])
v, _ = np.histogram(hsv_values[:, 2], bins=10)
hist_fea = np.concatenate((hs.reshape(100, ), v))
hsv_feats.append(hist_fea)
patch_colors.append(p)
return np.asarray(hsv_feats), patch_colors
def get_gt_label_info(fid_box, fid_obj):
""" Ground truth patch color (RGB) to label mapping """
gt_files = os.listdir(SEG_DIR)
gp_label = defaultdict(int)
for gt_file in gt_files:
if gt_file[-4:] != EXT:
continue
fid = gt_file.split(".")[0]
gt_file = SEG_DIR + gt_file
# 1. read the ground truth segmented file
gt_patches = get_patch_info(gt_file)
g_img = get_RGB(gt_file)
for gp in gt_patches:
if gp == CREAM:
continue
elif gp == BGR:
label = BGR_LABEL
else:
gt_pixels = np.where((g_img == gp).all(axis=2))
box = fid_box[fid]
labs = fid_obj[fid]
label = get_me_the_label(gt_pixels, box, labs)
if label is not None:
gp_label[gp] = label
return gp_label
def parallel_data_prep_svm(lst_fids):
""" Save label information for each patch at each scale level.
Parallel execution """
info_dir = FEAT_DIR + "patch_label_info/"
patch_subd = os.listdir(PATCH_DIR)
for fid in lst_fids:
patch_label_lst = []
for pd in patch_subd:
pfile = PATCH_DIR + pd + "/" + fid + EXT
p_part = "/".join(pfile.replace(EXT, "").split("/")[-2:])
p_img = get_RGB(pfile)
n_patches = get_patch_info(pfile)
for p in n_patches:
patch_pixels = np.where((p_img == p).all(axis=2))
# find out the label of this patch_pixels
# using manual segmentation info
patch_label = find_overlap_with_gt(patch_pixels, GP_LABEL, fid)
patch_label_lst.append([patch_label, p, p_part])
# print(patch_label, p, p_part)
pickle.dump(patch_label_lst, open(info_dir + fid + ".pkl", "wb"))
def find_overlap(p1, p2_info, s1_ifile, s2_ifile):
s1_img = get_RGB(s1_ifile)
s2_img = get_RGB(s2_ifile)
p1_pixels = np.where((s1_img == p1).all(axis=2))
p1_xy = [xy for xy in zip(p1_pixels[0], p1_pixels[1])]
max_cnt = 0
max_overlap_patch = None
for p2 in p2_info:
p2_pixels = np.where((s2_img == p2).all(axis=2))
p2_xy = [xy for xy in zip(p2_pixels[0], p2_pixels[1])]
cnt = len(set(p1_xy) & set(p2_xy))
if cnt > max_cnt:
max_overlap_patch = p2
max_cnt = cnt
return max_cnt, max_overlap_patch
def compute_patch_sift_hist(ifile, pfile, f_c_ixs):
""" Given img file, patched img file and corresponding VQ indices of the
keypoints, return histogram features """
def is_kp_in_patch(pp, kp_pt):
""" Check if key point is in patch """
flag = False
x_min = pp[0].min()
x_max = pp[0].max()
y_min = pp[1].min()
y_max = pp[1].max()
if kp_pt[0] >= x_min and kp_pt[0] <= x_max:
if kp_pt[1] >= y_min and kp_pt[1] <= y_max:
flag = True
return flag
fid = os.path.splitext(os.path.basename(ifile))[0]
kpts = pickle.load(open(FEAT_DIR + "kp/kp_" + fid + ".pkl", "rb"))
p_img = get_RGB(pfile)
pi = PIL.Image.open(pfile)
n_patches = defaultdict(int)
for pixel in pi.getdata():
n_patches[pixel] += 1
if n_patches is None:
print("\nNo patches:", pfile)
return None
hist_feats = []
patch_colors = []
for p in n_patches:
patch_pixels = np.where((p_img == p).all(axis=2))
patch_cixs = []
for kp_ix, kp in enumerate(kpts):
if is_kp_in_patch(patch_pixels, kp):
patch_cixs.append(f_c_ixs[kp_ix])
# print(patch_cixs)
if len(patch_cixs) > 0:
# print(patch_cixs)
h, _ = np.histogram(patch_cixs, bins=np.arange(0, 1001))
# print(h.shape)
# print(h)
# break
hist_feats.append(h)
patch_colors.append(p)
# break
hist_feats = np.asarray(hist_feats)
# print(hist_feats.shape, np.count_nonzero(hist_feats))
return hist_feats, patch_colors
def find_overlap_with_gt(pt_pixels, gp_label, fid):
""" Find overlap of unsupervised patch with the ground truth patch """
patch_label = BGR_LABEL
pt_xy = [(x, y) for x, y in zip(pt_pixels[0], pt_pixels[1])]
gt_file = SEG_DIR + fid + EXT
# 1. read the ground truth segmented file
gt_patches = get_patch_info(gt_file)
g_img = get_RGB(gt_file)
for i, gp in enumerate(gt_patches):
gt_pixels = np.where((g_img == gp).all(axis=2))
gt_xy = [xy for xy in zip(gt_pixels[0], gt_pixels[1])]
cnt = len(set(pt_xy) & set(gt_xy))
if cnt >= (len(pt_xy) * 0.75):
if gp == CREAM:
patch_label = None
elif gp == BGR:
patch_label = BGR_LABEL
else:
patch_label = gp_label[gp]
break # if > 75 overlap is found, return the label
else:
if gp == CREAM:
patch_label = BGR_LABEL
else:
patch_label = int(-1 * gp_label[gp])
return patch_label