def create_json_data(voc_path, output_path):
'''
Create json file saved data
voc_path: Path to VOC folder
output_path: Path to save folder
'''
#TRain
train_images = list()
train_objects = list()
with open(os.path.join(voc_path, "ImageSets/Main/trainval.txt")) as f:
ids = f.read().splitlines()
for id in ids:
#Get object, image in XML file
object_path = os.path.join(voc_path, "Annotations", id + ".xml")
objects = parse_annotation(object_path)
if len(objects) == 0:
continue
train_objects.append(objects)
image_path = voc_path + "/JPEGImages/"+id+".jpg"
train_images.append(image_path)
assert len(train_objects) == len(train_images)
#Save to file
with open(os.path.join(output_path, "TRAIN_images.json"), "w") as j:
json.dump(train_images, j)
with open(os.path.join(output_path, "TRAIN_objects.json"), "w") as j:
json.dump(train_objects, j)
#Test
test_images = list()
test_objects = list()
with open(os.path.join(voc_path, "ImageSets/Main/test.txt")) as f:
ids = f.read().splitlines()
for id in ids:
object_path = os.path.join(voc_path, "Annotations", id + ".xml")
objects = parse_annotation(object_path)
if len(objects) == 0:
continue
test_objects.append(objects)
test_image_path = voc_path + "/JPEGImages/"+id+".jpg"
test_images.append(test_image_path)
assert len(test_images) == len(test_objects)
#Save to file
with open(os.path.join(output_path, "TEST_images.json"), "w") as j:
json.dump(test_images, j)
with open(os.path.join(output_path, "TEST_objects.json"), "w") as j:
json.dump(test_objects, j)
def generate_anchors(ann_dir, img_dir, labels, net_input_size, num_anchors):
train_imgs, train_labels = parse_annotation(ann_dir, img_dir, labels)
grid_w = net_input_size / 32
grid_h = net_input_size / 32
# run k_mean to find the anchors
annotation_dims = []
for image in train_imgs:
cell_w = image['width'] / grid_w
cell_h = image['height'] / grid_h
for obj in image['object']:
relative_w = (float(obj['xmax']) - float(obj['xmin'])) / cell_w
relatice_h = (float(obj["ymax"]) - float(obj['ymin'])) / cell_h
annotation_dims.append(tuple(map(float, (relative_w, relatice_h))))
annotation_dims = np.array(annotation_dims)
centroids = run_kmeans(annotation_dims, num_anchors)
# write anchors to file
print('\naverage IOU for', num_anchors, 'anchors:',
'%0.2f' % avg_IOU(annotation_dims, centroids))
print_anchors(centroids)
def process(params):
input_txt_paths, output_txt_path = params
assert(len(input_txt_paths) > 0)
image_size_list = []
areas_list = []
for input_txt_path in input_txt_paths:
with open(input_txt_path, "r") as f:
lines = f.readlines()
image_size, areas = utils.parse_annotation(lines)
image_size_list.append(image_size)
areas_list.append(areas)
image_size = image_size_list[0]
for s in image_size_list:
assert(s == image_size)
areas = list(itertools.chain.from_iterable(areas_list))
num_areas = len(areas)
areas = [f"{x0} {y0} {x1} {y1} 1" for x0, y0, x1, y1 in areas]
areas = "\n".join(areas)
annotation = f"{image_size[0]} {image_size[1]}\n{num_areas}\n{areas}"
with open(output_txt_path, "w") as f:
f.write(annotation)
def process(params):
input_txt_path, input_png_path, output_png_path, rotation = params
with open(input_txt_path, "r") as f:
lines = f.readlines()
image_size, areas = utils.parse_annotation(lines)
image = utils.load_image_with_rotation(input_png_path, rotation)
image_h, image_w, _ = image.shape
assert (image_w == image_size[0] and image_h == image_size[1])
cv2.imwrite(str(output_png_path), image)
def process(params):
txt_filename, annotations = params
input_txt_path = input_dir / txt_filename
output_txt_path = output_dir / txt_filename
if not input_txt_path.exists():
print(input_txt_path.name, "does not exist, skipped.")
return None
with open(input_txt_path, "r") as f:
lines = f.readlines()
image_size, areas = utils.parse_annotation(lines)
image_w, image_h = image_size
if len(areas) <= 0:
return None
if len(areas) <= max([i for i, _ in annotations]):
print(target_file.name, " annotation is not valid, skipped.")
return None
areas = [(areas[i], s) for i, s in annotations]
areas_scaled = []
for (x0, y0, x1, y1), scale in areas:
#print(x0, y0, x1, y1)
orig_w, orig_h = (x1 - x0, y1 - y0)
assert (orig_h > 0 and orig_w > 0)
remove_ratio = 1.0 - (1.0 - scale_factor)**scale
trim_w = math.ceil(orig_w * remove_ratio * 0.5)
trim_h = math.ceil(orig_h * remove_ratio * 0.5)
x0 = min(x0 + trim_w, image_w - 1)
y0 = min(y0 + trim_h, image_h - 1)
x1 = max(x1 - trim_w, 0)
y1 = max(y1 - trim_h, 0)
#print(x0, y0, x1, y1)
assert (x0 < x1 and y0 < y1)
areas_scaled.append((x0, y0, x1, y1))
#break
areas = [f"{x0} {y0} {x1} {y1} 1" for x0, y0, x1, y1 in areas_scaled]
annotations = "\n".join(areas)
annotations = f"{image_size[0]} {image_size[1]}\n{len(areas)}\n{annotations}"
with open(output_txt_path, "w") as f:
f.write(annotations)
def process(params):
input_path, output_path = params
#print(input_path)
with open(input_path, "r") as f:
lines = f.readlines()
image_size, areas = utils.parse_annotation(lines)
input_areas = [(x0, y0, x1, y1, (x1 - x0) * (y1 - y0))
for x0, y0, x1, y1 in areas]
output_areas = []
for x0i, y0i, x1i, y1i, ai in input_areas:
need_to_append = True
for idx, (x0o, y0o, x1o, y1o, ao, no) in enumerate(output_areas):
x0u, y0u, x1u, y1u = (max(x0i, x0o), max(y0i, y0o), min(x1i, x1o),
min(y1i, y1o))
if (x1u - x0u <= 0 or y1u - y0u <= 0):
continue
au = (x1u - x0u) * (y1u - y0u)
iou = 2 * au / (ai + ao)
assert (iou <= 1)
if (iou >= iou_merge_threshold):
no_new = no + 1
x0 = max(0, int((x0o * no + x0i) / no_new))
y0 = max(0, int((y0o * no + y0i) / no_new))
x1 = min(image_size[0] - 1, int((x1o * no + x1i) / no_new))
y1 = min(image_size[1] - 1, int((y1o * no + y1i) / no_new))
assert (x1 - x0 > 0)
assert (y1 - y0 > 0)
a = (x1 - x0) * (y1 - y0)
output_areas[idx] = (x0, y0, x1, y1, a, no_new)
need_to_append = False
break
if need_to_append:
output_areas.append((x0i, y0i, x1i, y1i, ai, 1))
areas = [f"{x0} {y0} {x1} {y1} 1" for x0, y0, x1, y1, _, _ in output_areas]
num_areas = len(areas)
areas = "\n".join(areas)
annotation = f"{image_size[0]} {image_size[1]}\n{num_areas}\n{areas}"
#print(len(input_areas), num_areas)
with open(output_path, "w") as f:
f.write(annotation)
def process(params):
input_txt_path, input_png_path, rotation = params
output_png_filename_prefix = f"{input_png_path.stem}_{rotation}"
with open(input_txt_path, "r") as f:
lines = f.readlines()
image_size, areas = utils.parse_annotation(lines)
image = utils.load_image_with_rotation(input_png_path, rotation)
image_h, image_w, _ = image.shape
assert(image_w == image_size[0] and image_h == image_size[1])
if (False):
plt.imshow(cv2.cvtColor(image, cv2.COLOR_BGR2RGB))
plt.show()
area_size_list = []
for i, area in enumerate(areas):
x0, y0, x1, y1 = area
crop_image = image[y0:y1, x0:x1, :]
crop_h, crop_w, _ = crop_image.shape
if (False):
plt.imshow(cv2.cvtColor(crop_image, cv2.COLOR_BGR2RGB))
plt.show()
if (True):
crop_image = cv2.resize(crop_image, (output_image_size, output_image_size), interpolation = cv2.INTER_CUBIC)
file_path = output_dir / f"{output_png_filename_prefix}_{i}.png"
cv2.imwrite(str(file_path), crop_image)
area_size_list.append((crop_h, crop_w))
return area_size_list
import json
import os
import numpy as np
from tqdm import tqdm
import re
from utils import parse_annotation
parser = argparse.ArgumentParser()
parser.add_argument('--data_dir',
default=os.path.join('/Users/UnicornKing/20180101_120040'))
args = parser.parse_args()
if __name__ == '__main__':
annotations = [(file, parse_annotation(os.path.join(args.data_dir, file)))
for file in sorted(os.listdir(args.data_dir))
if os.path.splitext(file)[-1] == '.json']
prev_file, prev_annot = annotations[0]
for file, annot in tqdm(annotations[1:]):
prev_labels = set([shape['label'] for shape in prev_annot['shapes']])
labels = set([shape['label'] for shape in annot['shapes']])
for label in prev_labels.intersection(labels):
if re.search(r'_\d+', label) is None:
continue
prev_points = [
shape['points'] for shape in prev_annot['shapes']
if shape['label'] == label
][0]
points = [
shape['points'] for shape in annot['shapes']
b = np.random.randint(0, 255, (len(action_to_idx), 1))
colors = np.column_stack([r, g, b]).tolist()
annot_cache = {}
video_writer = cv2.VideoWriter('out.avi', 0, args.fps, (1200, 675))
for i, frame in tqdm(
enumerate(
sorted(glob.glob(os.path.join(args.frames_dir,
'*.jpg')))[:25000][::2])):
frame_name = os.path.splitext(frame)[0]
annot_name = frame_name + '.json'
annot_path = os.path.join(args.frames_dir, annot_name)
image_path = os.path.join(args.frames_dir, frame)
image = cv2.imread(image_path)
if os.path.exists(annot_path):
annot = parse_annotation(annot_path)
for shape in annot['shapes']:
label = shape['label']
if re.sub(r'_\d+', '', label) not in action_to_idx:
continue
annot_cache.setdefault(label, [])
annot_cache[label].append({
'image': image,
'points': shape['points']
})
if len(annot_cache[label]) == args.num_segments:
if label not in meta['action'].values:
from utils import parse_annotation, make_dir_if_needed
import cv2
import re
import numpy as np
from tqdm import tqdm
from definitions import action_to_idx, idx_to_action
parser = argparse.ArgumentParser()
parser.add_argument('--data_dir', default='/Users/UnicornKing/20180101_120040')
parser.add_argument('--output_image_dir', default='data/actions')
parser.add_argument('--output_label_path', default='data/actions.txt')
parser.add_argument('--split_length', type=int, default=60)
args = parser.parse_args()
if __name__ == '__main__':
annotations = [(file, parse_annotation(os.path.join(args.data_dir, file))) for file in
sorted(os.listdir(args.data_dir))
if os.path.splitext(file)[-1] == '.json']
label_counts = {}
splits_counts = {}
records = []
dirs = set()
for file, annot in tqdm(annotations):
file_name = os.path.splitext(file)[0]
for i, shape in enumerate(annot['shapes']):
x1, y1 = shape['points'][0]
x2, y2 = shape['points'][1]
label = shape['label']
if re.sub(r'_\d+', '', label) not in action_to_idx:
continue
from keras.callbacks import EarlyStopping, ModelCheckpoint
from keras.optimizers import SGD, Adagrad
import numpy as np
import tensorflow as tf
from parameter import *
from utils import parse_annotation, data_gen
from Model2 import model
all_img = parse_annotation(ann_dir)
def custom_loss(y_true, y_pred):
### Adjust prediction
# adjust x and y
pred_box_xy = tf.sigmoid(y_pred[:, :, :, :, :2])
# adjust w and h
pred_box_wh = tf.exp(y_pred[:, :, :, :, 2:4]) * np.reshape(
ANCHORS, [1, 1, 1, BOX, 2])
pred_box_wh = tf.sqrt(
pred_box_wh /
np.reshape([float(GRID_W), float(GRID_H)], [1, 1, 1, 1, 2]))
# adjust confidence
pred_box_conf = tf.expand_dims(tf.sigmoid(y_pred[:, :, :, :, 4]), -1)
# adjust probability
pred_box_prob = tf.nn.softmax(y_pred[:, :, :, :, 5:])
y_pred = tf.concat(