def visualize_dt_and_gt(images, y_true, y_pred, config):
"""Takes a batch of images and creates bounding box visualization on top
Arguments:
images {[type]} -- numpy tensor of images
y_true {[type]} -- tensor of ground truth
y_pred {[type]} -- tensor of predictions
config {[type]} -- dict of various hyperparameters
Returns:
[type] -- dict of various hyperparameters
"""
img_with_boxes = []
all_filtered_boxes, all_filtered_classes, all_filtered_scores = filter_batch(
y_pred, config)
box_input = y_true[:, :, 1:5]
labels = y_true[:, :, 9:]
font = cv2.FONT_HERSHEY_SIMPLEX
for i, img in enumerate(images):
non_zero_boxes = box_input[i][box_input[i] > 0].reshape((-1, 4))
non_zero_labels = []
for k, coords in enumerate(box_input[i, :]):
if np.sum(coords) > 0:
for j, l in enumerate(labels[i, k]):
if l == 1:
non_zero_labels.append(j)
for j, det_box in enumerate(all_filtered_boxes[i]):
det_box = bbox_transform_single_box(det_box)
cv2.rectangle(img, (det_box[0], det_box[1]),
(det_box[2], det_box[3]), (0, 0, 255), 1)
cv2.putText(img, config.CLASS_NAMES[all_filtered_classes[i][j]] + " " + str(
all_filtered_scores[i][j]), (det_box[0], det_box[1]), font, 0.5, (0, 0, 255), 1, cv2.LINE_AA)
for j, gt_box in enumerate(non_zero_boxes):
gt_box = bbox_transform_single_box(gt_box)
cv2.rectangle(img, (gt_box[0], gt_box[1]),
(gt_box[2], gt_box[3]), (0, 255, 0), 1)
cv2.putText(img, config.CLASS_NAMES[int(non_zero_labels[j])], (gt_box[0], gt_box[1]), font, 0.5,
(0, 255, 0), 1, cv2.LINE_AA)
img_with_boxes.append(img[:, :, [2, 1, 0]])
return img_with_boxes
def eval(img_file):
#create config object
config = load_dict(CONFIG)
config.BATCH_SIZE = 1 #文件数量
config.FINAL_THRESHOLD = 0.5 #conf阈值
'''
#open files with images and ground truths files with full path names
with open(imgs_list) as imgs:
img_names = imgs.read().splitlines()
imgs.close()
'''
#hide the other gpus so tensorflow only uses this one
os.environ['CUDA_VISIBLE_DEVICES'] = CUDA_VISIBLE_DEVICES
#tf config and session
cfg = tf.ConfigProto(allow_soft_placement=True)
sess = tf.Session(config=cfg)
K.set_session(sess)
#instantiate model
squeeze = SqueezeDet(config)
squeeze.model.load_weights(checkpoint_dir)
img = cv2.imread(img_file).astype(np.float32, copy=False)
orig_h, orig_w, _ = [float(v) for v in img.shape]
# scale image
draw_img = img
img = cv2.resize(img, (config.IMAGE_WIDTH, config.IMAGE_HEIGHT))
img = (img - np.mean(img)) / np.std(img)
img = np.reshape(img, (1, config.IMAGE_HEIGHT, config.IMAGE_WIDTH, 3))
y_pred = squeeze.model.predict(img)
#filter batch with nms
all_filtered_boxes, all_filtered_classes, all_filtered_scores = filter_batch(
y_pred, config)
font = cv2.FONT_HERSHEY_SIMPLEX
for j, det_box in enumerate(all_filtered_boxes[0]):
#transform into xmin, ymin, xmax, ymax
det_box = bbox_transform_single_box(det_box)
print(all_filtered_scores[0][j])
#add rectangle and text
cv2.rectangle(draw_img, (det_box[0], det_box[1]),
(det_box[2], det_box[3]), (0, 0, 255), 1)
cv2.putText(draw_img, 'head' + str(all_filtered_scores[0][j]),
(det_box[0], det_box[1]), font, 0.5, (0, 255, 0), 1,
cv2.LINE_AA)
cv2.imwrite('restul.png', draw_img)
def visualize_dt_and_gt(images, y_true, y_pred, config):
"""Takes a batch of images and creates bounding box visualization on top
Arguments:
images {[type]} -- numpy tensor of images
y_true {[type]} -- tensor of ground truth
y_pred {[type]} -- tensor of predictions
config {[type]} -- dict of various hyperparameters
Returns:
[type] -- dict of various hyperparameters
"""
#normalize for printing
#images = (images - np.min(images, axis=(1,2,3), keepdims=1)) / np.max(images, axis=(1,2,3), keepdims=1) - np.min(images, axis=(1,2,3), keepdims=1)
img_with_boxes = []
#filter batch with nms
all_filtered_boxes, all_filtered_classes, all_filtered_scores = filter_batch(
y_pred, config)
#print(len(all_filtered_scores))
#print(len(all_filtered_scores[0]))
#get gt boxes
box_input = y_true[:, :, 1:5]
#and gt labels
labels = y_true[:, :, 9:]
font = cv2.FONT_HERSHEY_SIMPLEX
#iterate images
for i, img in enumerate(images):
#get predicted boxes
non_zero_boxes = box_input[i][box_input[i] > 0].reshape((-1, 4))
non_zero_labels = []
img = cv2.resize(img, (1280, 1080)) #!!!!!
#get the predicted labels
for k, coords in enumerate(box_input[i, :]):
if np.sum(coords) > 0:
for j, l in enumerate(labels[i, k]):
if l == 1:
non_zero_labels.append(j)
#iterate predicted boxes
for j, det_box in enumerate(all_filtered_boxes[i]):
#transform into xmin, ymin, xmax, ymax
det_box = bbox_transform_single_box(det_box)
x_scale = 1280 / config.IMAGE_WIDTH
y_scale = 1080 / config.IMAGE_HEIGHT
det_box[0], det_box[2] = int(det_box[0] * x_scale), int(
det_box[2] * x_scale)
det_box[1], det_box[3] = int(det_box[1] * y_scale), int(
det_box[3] * y_scale)
#add rectangle and text
cv2.rectangle(img, (det_box[0], det_box[1]),
(det_box[2], det_box[3]), (255, 10, 10), 2)
cv2.putText(
img, config.CLASS_NAMES[all_filtered_classes[i][j]] + " " +
str(all_filtered_scores[i][j]), (det_box[0], det_box[1]), font,
1, (255, 10, 10), 2, cv2.LINE_AA)
#iterate gt boxes
for j, gt_box in enumerate(non_zero_boxes):
#transform into xmin, ymin, xmax, ymax
gt_box = bbox_transform_single_box(gt_box)
x_scale = 1280 / config.IMAGE_WIDTH
y_scale = 1080 / config.IMAGE_HEIGHT
gt_box[0], gt_box[2] = int(gt_box[0] * x_scale), int(gt_box[2] *
x_scale)
gt_box[1], gt_box[3] = int(gt_box[1] * y_scale), int(gt_box[3] *
y_scale)
#add rectangle and text
cv2.rectangle(img, (gt_box[0], gt_box[1]), (gt_box[2], gt_box[3]),
(0, 255, 0), 2)
cv2.putText(img, config.CLASS_NAMES[int(non_zero_labels[j])],
(gt_box[0], gt_box[1]), font, 1, (0, 255, 0), 1,
cv2.LINE_AA)
#chagne to
#img_with_boxes.append(img[:,:, [2,1,0]]) #rgb
img_with_boxes.append(img[:, :, [0, 1, 2]]) #bgr
return img_with_boxes
img_name = "sm1.jpg"
# open img
img = cv2.imread(img_name).astype(np.float32, copy=False)
orig_h =img.shape[0]
orig_w =img.shape[1]
# subtract means
img = (img - np.mean(img)) / np.std(img)
img = cv2.resize(img, (cfg.IMAGE_WIDTH, cfg.IMAGE_HEIGHT))
img_for_pred = np.expand_dims(img, axis=0)
print (img_for_pred.shape)
y_pred = model.predict(img_for_pred)
print(" predicted something")
all_filtered_boxes, all_filtered_classes, all_filtered_scores = filter_batch(y_pred, cfg)
font = cv2.FONT_HERSHEY_SIMPLEX
img = cv2.imread(img_name)
for j, det_box in enumerate(all_filtered_boxes[0]):
print det_box
"""
# transform into xmin, ymin, xmax, ymax
det_box = bbox_transform_single_box(det_box)
x_scale = orig_w / cfg.IMAGE_WIDTH
y_scale = orig_h / cfg.IMAGE_HEIGHT
det_box[0], det_box[2] = int(det_box[0] * x_scale), int(det_box[2] * x_scale)
det_box[1], det_box[3] = int(det_box[1] * y_scale), int(det_box[3] * y_scale)
# add rectangle and text