def load_image(self, image_index):
path = os.path.join(self.images_path, self.image_names[image_index] + '.jpg')
return read_image_bgr(path)
# test_img_list = test_img_list[:1000]
print('数据集{}总数：{}'.format(current_config.voc_sub_dir, len(test_img_list)))
for id, imgf in enumerate(test_img_list):
# imgfp = os.path.join(config.test_images_path, imgf)
imgfp = imgf['filepath']
# if test_img_list[id]['filename'] == '000227.jpg':
# print(id, test_img_list[id])
# exit()
if os.path.isfile(imgfp):
try:
img = read_image_bgr(imgfp)
except:
continue
img = preprocess_image(img.copy())
img, scale = resize_image(img,
min_side=config.img_min_size,
max_side=config.img_max_size)
orig_image = read_image_rgb(imgfp)
_, _, detections = model.predict_on_batch(np.expand_dims(img, axis=0))
detections[:, :, 0] = np.maximum(0, detections[:, :, 0])
detections[:, :, 1] = np.maximum(0, detections[:, :, 1])
detections[:, :, 2] = np.minimum(img.shape[1], detections[:, :, 2])
def inference():
set_runtime_environment()
start_time = time.time()
config = Config('configRetinaNet.json')
wpath = config.trained_weights_path
result_path = config.test_result_path
txt_path = result_path.replace('results', 'txt')
classes = config.classes
if not os.path.exists(result_path):
os.makedirs(result_path)
if not os.path.exists(txt_path):
os.makedirs(txt_path)
model, _ = resnet_retinanet(len(classes),
backbone=config.type,
weights='imagenet',
nms=True,
config=config)
model.load_weights(wpath, by_name=True, skip_mismatch=True)
files = sorted(os.listdir(config.test_images_path))
for nimage, imgf in enumerate(files):
# if imgf not in ['ship0201606110201801.jpg', 'ship0201606110201902.jpg', 'ship02016061102012014.jpg', 'ship0201711030202902.jpg']:
# continue
# if nimage >= int(len(files) * 0.1):
# break
imgfp = os.path.join(config.test_images_path, imgf)
if os.path.isfile(imgfp):
try:
img = read_image_bgr(imgfp)
except:
continue
img = preprocess_image(img.copy())
img, scale = resize_image(img,
min_side=config.img_min_size,
max_side=config.img_max_size)
_, _, detections = model.predict_on_batch(
np.expand_dims(img, axis=0))
# bbox要取到边界内
detections[:, :, 0] = np.maximum(0, detections[:, :, 0])
detections[:, :, 1] = np.maximum(0, detections[:, :, 1])
detections[:, :, 2] = np.minimum(img.shape[1], detections[:, :, 2])
detections[:, :, 3] = np.minimum(img.shape[0], detections[:, :, 3])
detections[0, :, :4] /= scale
scores = detections[0, :, 4:]
# 推测置信度
indices = np.where(detections[0, :, 4:] >= 0.05)
scores = scores[indices]
scores_sort = np.argsort(-scores)[:100]
image_boxes = detections[0, indices[0][scores_sort], :4]
image_scores = np.expand_dims(
detections[0, indices[0][scores_sort],
4 + indices[1][scores_sort]],
axis=1)
image_detections = np.append(image_boxes, image_scores, axis=1)
image_predicted_labels = indices[1][scores_sort]
txtfile = imgf.replace('.jpg', '.txt')
realpath = os.path.join(txt_path, txtfile)
f = open(realpath, 'w', encoding='utf-8')
if len(image_boxes) > 0:
for i, box in enumerate(image_boxes):
xmin = int(box[0])
ymin = int(box[1])
xmax = int(box[2])
ymax = int(box[3])
# print(xmin, ymin, xmax, ymax)
f.write('{} {} {} {} {} {}\n'.format(
classes[image_predicted_labels[i]], xmin, ymin, xmax,
ymax, image_scores[i][0]))
f.close()
print("生成txt '" + txtfile + "'" +
' time:{}, 目标框：{}'.format(time.time() -
start_time, len(image_boxes)))
start_time = time.time()