def evaluation_gt():
#read labels path
gt_path = '/home/alupotto/data/autel/new_labels'
gt_list = glob.glob("{}/*.txt".format(gt_path))
#declare dictionary gt
classes = load_classes('data/autel.names')
dict_gt = dict.fromkeys(classes, 0)
for gt_file in gt_list:
with open(gt_file) as f_gt:
gt_lines = f_gt.readlines()
#clean \n & convert to a list
gt_lines = [line.strip('\n').split(' ') for line in gt_lines]
#convert to string to float
gt_lines = [int(gt_line[0]) for gt_line in gt_lines]
for class_id in gt_lines:
class_name = classes[class_id]
if class_name in dict_gt:
dict_gt[class_name] += 1
else:
dict_gt[class_name] = 1
with open('pickles/dict_gt.pkl', 'wb') as gt_pkl:
pickle.dump(dict_gt, gt_pkl, protocol=pickle.HIGHEST_PROTOCOL)
return dict_gt
def check_resolutions(imgs_path):
#file_resolutions = open('files_check/resolution.txt', 'w')
#load classes autel
classes = load_classes('data/autel.resolutions')
#dictionary count resolution
dict_res_img = dict.fromkeys(classes, 0)
dict_videos = dict.fromkeys(classes)
dict_res_videos = {k: [] for k in dict_videos}
#load images
img_files = sorted(glob.glob('%s**/*.jpg' % imgs_path))
#load labels
label_files = sorted(glob.glob('%s**/*.xml' % imgs_path))
#img with bad resolution frames
folder_id = imgs_path.split('/')[-2]
#img_bad_res = open("../files_check/frames_bad_{}.txt".format(folder_id), 'w')
for img_path in img_files:
#check if image has label
if img_path.replace('.jpg', '.xml') in label_files:
#video id
video_id = img_path.split('/')[-1].split('_')[-2]
#extract image shape
img = cv2.imread(img_path)
h_img, w_img, _ = img.shape
#class id
cls_id = resolution_id(w_img, h_img)
class_name = classes[cls_id]
if not video_id in dict_res_videos[class_name]:
dict_res_videos[class_name].append(video_id)
#num images dictionary
if class_name in dict_res_img:
dict_res_img[class_name] += 1
else:
dict_res_img[class_name] = 1
#pickle resolution images
with open('pickles/dict_res_img_{}.pkl'.format(folder_id),
'wb') as res_pkl:
pickle.dump(dict_res_img, res_pkl)
#pickle resolution videos
with open('pickles/dict_res_videos_{}.pkl'.format(folder_id),
'wb') as res_pkl:
pickle.dump(dict_res_videos, res_pkl)
return dict_res_img, dict_res_videos
def create_labels_yolo(images_path, labels_path):
#load classes autel
classes = load_classes('data/autel.names')
#load images
img_files = sorted(glob.glob('%s/*.jpg' % images_path))
#load labels
label_files = sorted(glob.glob('%s/*.xml' % images_path))
for img_path in img_files:
video_id = img_path.split('/')[-1].split('_')[-2]
if img_path.replace('.jpg', '.xml') in label_files:
# xml path
xml_path = img_path.replace('.jpg', '.xml')
# root xml
root = ET.parse(xml_path)
# check real image shape with shape of the label
h_xml = int(root.find('size').find('height').text)
w_xml = int(root.find('size').find('width').text)
# extract image shape
img = cv2.imread(img_path)
h_img, w_img, _ = img.shape
#check resolution objects
if check_resolution(h_img, w_img, video_id):
#label file generation
label_id = img_path.replace('.jpg', '.txt').split('/')[-1]
#prompt
print(labels_path + label_id)
#file label
gt_file = open(labels_path + label_id, 'w')
for obj in root.findall('object'):
obj_name = obj.find('name').text
if obj_name in classes:
#parse xml
bndbox = obj.find('bndbox')
x0 = float(bndbox.find('xmin').text)
y0 = float(bndbox.find('ymin').text)
x1 = float(bndbox.find('xmax').text)
y1 = float(bndbox.find('ymax').text)
#prepare coords for conversion
obj_id = classes.index(obj_name)
bbox_xml = (x0, x1, y0, y1)
#print in txt file
gt_file.write(
str(obj_id) + " " +
" ".join([str(coord)
for coord in bbox_xml]) + '\n')
def check_labels(imgs_path):
#load classes autel
classes = load_classes('data/autel.names')
#dictionary count classes
dict_gt = dict.fromkeys(classes, 0)
#dictionary count resolution
dict_resolution = {}
#load images
img_files = sorted(glob.glob('%s/*.jpg' % imgs_path))
#load labels
label_files = sorted(glob.glob('%s/*.xml' % imgs_path))
#img with bad resolution frames
folder_id = imgs_path.split('/')[-2]
img_bad_res = open("files_check/frames_bad_{}.txt".format(folder_id), 'w')
for img_path in img_files:
# print(dict_gt)
#check if image has label
if img_path.replace('.jpg', '.xml') in label_files:
#xml path
xml_path = img_path.replace('.jpg', '.xml')
#root xml
root = ET.parse(xml_path)
#check real image shape with shape of the label
h_xml = int(root.find('size').find('height').text)
w_xml = int(root.find('size').find('width').text)
#extract image shape
img = cv2.imread(img_path)
h_img, w_img, ch_img = img.shape
#check xml & jpg are equal
if check_size_objects((h_img, w_img), (h_xml, w_xml), root):
for obj in root.findall('object'):
obj_name = obj.find('name').text
#count objects
if obj_name in dict_gt:
dict_gt[obj_name] += 1
else:
dict_gt[obj_name] = 1
with open('pickles/dict_{}.pkl'.format(folder_id), 'wb') as gt_pkl:
pickle.dump(dict_gt, gt_pkl)
def histogram_resolutions_img(dict_res_img=None):
if dict_res_img is None:
with open('pickles/dict_res_img_images.pkl', 'rb') as gt_pkl:
dict_res_img = pickle.load(gt_pkl)
classes = load_classes('data/autel.resolutions')
window_title = "img/resolution"
plot_title = "num images per resolution"
x_label = "num images"
output_path = "output/resolution_img_08072018.png"
to_show = True
plot_color = 'royalblue'
draw_plot_func(dict_res_img, len(classes), window_title, plot_title,
x_label, output_path, to_show, plot_color, "")
def check_devset():
# load classes with background
classes = load_classes('data/autel_background.names')
with open('/home/alupotto/data/autel_08102018/autel_devset.txt',
'r') as file_dev:
list_paths = file_dev.readlines()
lbls_path = [
img_path.replace('images', 'labels_yolo_withbackground').replace(
'.jpg', '.txt').strip('\n') for img_path in list_paths
]
dev_dict = count_classes(lbls_path, classes)
print(dev_dict)
def histogram_resolutions_videos(dict_res_video=None):
if dict_res_video is None:
with open('pickles/dict_res_videos_images.pkl', 'rb') as gt_pkl:
dict_res_video = pickle.load(gt_pkl)
for key, values in dict_res_video.items():
dict_res_video[key] = len(values)
classes = load_classes('data/autel.resolutions')
window_title = "ground truth"
plot_title = "number videos per resolution"
x_label = "num videos"
output_path = "output/resolution_videos_08072018.png"
to_show = True
plot_color = 'royalblue'
draw_plot_func(dict_res_video, len(classes), window_title, plot_title,
x_label, output_path, to_show, plot_color, "")
def histogram_classes_gt(dict_gt=None):
if dict_gt is None:
with open('pickle/dict_gt.pkl', 'rb') as gt_pkl:
dict_gt = pickle.load(gt_pkl)
classes = load_classes('data/autel.names')
gt_path = '/home/alupotto/data/autel/new_labels'
gt_list = glob.glob("{}/*.txt".format(gt_path))
window_title = "ground truth"
plot_title = "Ground-Truth\n"
plot_title += "(" + str(len(gt_list)) + " files and " + str(
len(classes)) + " classes)"
x_label = "number ground truth objects"
output_path = "output/ground_truth.png"
to_show = False
plot_color = 'forestgreen'
draw_plot_func(dict_gt, len(classes), window_title, plot_title, x_label,
output_path, to_show, plot_color, "")
return dict_gt
def histogram_classes_AP(dict_classes=None):
if dict_classes is None:
with open('pickles/dict_classes.pkl', 'rb') as cls_pkl:
dict_classes = pickle.load(cls_pkl)
classes = load_classes('data/autel.names')
for key, value in dict_classes.items():
dict_classes[key] = np.mean(value)
mAP = dict_classes['mAP']
dict_classes.pop('mAP')
window_title = "AP per class"
plot_title = "mAP = {0:.2f}%".format(mAP * 100)
x_label = "Average Precision"
output_path = "output/classes_AP.png"
to_show = True
plot_color = 'royalblue'
draw_plot_func(dict_classes, len(classes), window_title, plot_title,
x_label, output_path, to_show, plot_color, "")
return dict_classes
def histogram_classes_gt(imgs_path, dict_gt=None):
if dict_gt is None:
with open('pickles/dict_part1_autel.pkl', 'rb') as gt_pkl:
dict_gt = pickle.load(gt_pkl)
print(dict_gt)
classes = load_classes('data/autel.names')
# load images
img_files = sorted(glob.glob('%s/*.jpg' % imgs_path))
print(len(img_files))
window_title = "ground truth"
plot_title = "Ground-Truth\n"
plot_title += "(" + str(len(img_files)) + " files and " + str(
len(classes)) + " classes)"
x_label = "number ground truth objects"
output_path = "output/ground_truth_08072018.png"
to_show = True
plot_color = 'forestgreen'
draw_plot_func(dict_gt, len(classes), window_title, plot_title, x_label,
output_path, to_show, plot_color, "")
def create_dev_set(labels_path, per_cent):
#load classes with background
classes = load_classes('data/autel_background.names')
#file_paths
file_paths = open(
os.path.join(os.path.dirname(labels_path.rstrip('/')),
'autel_devset.txt'), 'w')
#load labels
label_files = sorted(glob.glob('%s/*.txt' % labels_path))
#check dict_classes
if not os.path.isfile('pickles/dict_gt.pkl'):
dict_classes = count_classes(label_files, classes)
else:
with open('pickles/dict_gt.pkl', 'rb') as gt_pkl:
dict_classes = pickle.load(gt_pkl)
#apply per_cent to classes in dict
for key, value in dict_classes.items():
dict_classes[key] = int(value * per_cent)
#shuffling
seed(23)
shuffle(label_files)
check_dict = {}
for label in label_files:
with open(label, 'r') as lbl_file:
file = lbl_file.readlines()
id_classes = [int(line.split(' ')[0]) for line in file]
for cls in id_classes:
dict_classes[classes[cls]] -= 1
if all(value > 0 for value in dict_classes.values()):
img_path = label.replace('labels_yolo_withbackground',
'images').replace('.txt', '.jpg')
file_paths.write(img_path + '\n')
def evaluation_classes(gt_path, pred_path):
MIN_OVERLAP = 0.5
ROOT_PATH = '/home/alupotto/data/autel/'
#AP: 0.1667(0.0556)
#AP: 0.85
#read labels path
#gt_path = '/home/alupotto/data/autel/new_labels'
gt_list = glob.glob("{}/*.txt".format(gt_path))
gt_list.sort()
#paths predicted
#pred_path = '/home/alupotto/data/autel/new_predicted'
#pred_path = '/home/alupotto/data/autel/experiments/predicted_2420'
APs = []
#declare dictionary classes
classes = load_classes('data/autel.names')
AP_classes = dict.fromkeys(classes)
AP_classes = {k: [] for k in AP_classes}
for idx, gt_file in enumerate(gt_list):
#read images path
with open(gt_file) as f_gt:
gt_lines = f_gt.readlines()
#clean \n & convert to a list
gt_lines = [line.strip('\n').split(' ') for line in gt_lines]
#convert to string to float
gt_lines = [list(map(float, gt_line)) for gt_line in gt_lines]
#extract gt_id
gt_id = gt_file.split('/')[-1]
pred_file = os.path.join(pred_path, gt_id)
#read prediction
with open(pred_file) as f_pred:
f_pred = f_pred.readlines()
# clean \n and covnert to a list
pred_lines = [line.strip('\n').split(' ') for line in f_pred]
#convert string to float
pred_lines = [list(map(float, pred_line)) for pred_line in pred_lines]
gt_lines = np.array(gt_lines)
bbox_gt = gt_lines[:, 1:]
AP_temp = {}
correct = []
detected = []
if len(pred_lines) == 0:
if len(gt_lines) != 0:
tpos_non = len(gt_lines) * [0]
for idx, (cls_gt, _, _, _, _) in enumerate(gt_lines):
cls_name = classes[int(cls_gt)]
tpos_non[idx] = int(0)
APs.append(0)
AP_temp[cls_name] = tpos_non
continue
tpos = len(pred_lines) * [0]
for idx, (cls_pred, conf, x1, y1, x2, y2) in enumerate(pred_lines):
bbox_pred = np.array((x1, y1, x2, y2))
bbox_pred = bbox_pred.reshape((1, bbox_pred.shape[0]))
iou = bbox_iou(bbox_pred, bbox_gt)
# extract index of largest overlap
best_i = np.argmax(iou)
cls_name = classes[int(cls_pred)]
if iou[best_i] > MIN_OVERLAP and cls_pred == gt_lines[
best_i, 0] and best_i not in detected:
correct.append(1)
detected.append(best_i)
tpos[idx] = int(1)
AP_temp[cls_name] = tpos
else:
correct.append(0)
tpos[idx] = int(0)
AP_temp[cls_name] = tpos
for cls, value in AP_temp.items():
true_positives = np.array(value)
false_positives = 1 - true_positives
false_positives = np.cumsum(false_positives)
true_positives = np.cumsum(true_positives)
recall = true_positives / len(gt_lines) if len(
gt_lines) else true_positives
precision = true_positives / np.maximum(
true_positives + false_positives,
np.finfo(np.float64).eps)
AP = compute_ap(recall, precision)
APs.append(AP)
if cls in AP_classes:
AP_classes[cls].append(AP)
else:
AP_classes[cls] = AP
print("Sample [%d/%d] AP: %.4f mAP: %.4f" %
(len(APs), len(gt_list), AP, np.mean(APs)))
AP_classes['mAP'] = np.mean(APs)
with open('pickles/dict_classes.pkl', 'wb') as cls_pkl:
pickle.dump(AP_classes, cls_pkl, protocol=pickle.HIGHEST_PROTOCOL)
return AP_classes
import cv2
config_path = 'config/yolov3.cfg'
weights_path = 'config/yolov3.weights'
class_path = 'config/coco.names'
img_size = 416
conf_thres = 0.8
nms_thres = 0.4
# Load model and weights
model = Darknet(config_path, img_size=img_size)
model.load_weights(weights_path)
model.cuda()
model.eval()
classes = load_classes(class_path)
Tensor = torch.cuda.FloatTensor
def detect_image(img):
# Scale and pad image
ratio = min(img_size / img.size[0], img_size / img.size[1])
imw = round(img.size[0] * ratio)
imh = round(img.size[1] * ratio)
img_transforms = transforms.Compose([
transforms.Resize((imh, imw)),
transforms.Pad((max(int(
(imh - imw) / 2), 0), max(int(
(imw - imh) / 2), 0), max(int(
(imh - imw) / 2), 0), max(int((imw - imh) / 2), 0)),
(128, 128, 128)),
parser.add_argument("--epochs", type=int, default=20, help="number of epochs")
parser.add_argument("--batch_size", type=int, default=16, help="size of each image batch")
parser.add_argument("--n_cpu", type=int, default=0, help="number of cpu threads used during batch generation")
parser.add_argument("--is_cuda", type=bool, default=True, help="whether to use cuda")
parser.add_argument("--class_path", type=str, default="data/fish.names", help="path to class label file")
parser.add_argument("--data_config_path", type=str, default="data/fish.data", help="path to data config file")
parser.add_argument("--model_config_path", type=str, default="config/fish.cfg", help="path to model config file")
parser.add_argument("--weights_path", type=str, default="config/yolov3.weights", help="path to pre-trained weight file")
parser.add_argument("--checkpoint_interval", type=int, default=1, help="interval between saving model weights")
parser.add_argument("--checkpoint_dir", type=str, default="checkpoints", help="directory where model checkpoints are saved")
args = parser.parse_args()
#print(args)
# Load data class and data configuration.
classes = load_classes(args.class_path)
data_cfg = parse_data_config(args.data_config_path)
train_path = data_cfg["train"]
# Load hyperparameters.
hyperparams = parse_model_config(args.model_config_path)[0]
learning_rate = float(hyperparams["learning_rate"])
momentum = float(hyperparams["momentum"])
decay = float(hyperparams["decay"])
burn_in = int(hyperparams["burn_in"])
###################################################################################3
# Create and initial model.
model = Darknet(args.model_config_path)
