def __disp_image(self, img, classes, bboxes):
bvalid = (classes !=0)
classes = classes[bvalid]
bboxes = bboxes[bvalid]
scores =np.full(classes.shape, 1.0)
visualization.plt_bboxes(img, classes, scores, bboxes,title='Ground Truth')
return
def run(self):
self.dataset_name = 'pascalvoc_2007'
self.dataset_split_name = 'train'
self.dataset_dir = '../../data/voc/tfrecords/'
self.batch_size = 32
with tf.Graph().as_default():
tensors = self.__get_images_labels_bboxes()
with tf.Session('') as sess:
init = tf.global_variables_initializer()
sess.run(init)
with slim.queues.QueueRunners(sess):
for i in range(5000):
image, shape, format, filename, glabels, gbboxes,gdifficults = sess.run(list(tensors))
if str(filename,'utf-8') == "000394":
print(str(filename,'utf-8'))
scores = np.full(glabels.shape, 1.0)
visualization.plt_bboxes(image, glabels, scores, gbboxes,title=filename)
plt.show()
break
return
def __disp_matched_anchors(self, img, target_labels_data,
target_localizations_data, target_scores_data):
found_matched = False
all_anchors = g_ssd_model.get_allanchors()
for i, target_score_data in enumerate(target_scores_data):
num_pos = (target_score_data > 0.5).sum()
if (num_pos == 0):
continue
print('Found {} matched default boxes in layer {}'.format(
num_pos, g_ssd_model.feat_layers[i]))
pos_sample_inds = (target_score_data > 0.5).nonzero()
pos_sample_inds = [
pos_sample_inds[0], pos_sample_inds[1], pos_sample_inds[2]
]
classes = target_labels_data[i][pos_sample_inds]
scores = target_scores_data[i][pos_sample_inds]
bboxes_default = g_ssd_model.get_allanchors(
minmaxformat=True)[i][pos_sample_inds]
bboxes_gt = g_ssd_model.decode_bboxes_layer(
target_localizations_data[i][pos_sample_inds],
all_anchors[i][pos_sample_inds])
print("default box minimum, {} gt box minimum, {}".format(
bboxes_default.min(), bboxes_gt.min()))
marks_default = np.full(classes.shape, True)
marks_gt = np.full(classes.shape, False)
scores_gt = np.full(scores.shape, 1.0)
bboxes = bboxes_default
neg_marks = marks_default
add_gt = True
if add_gt:
bboxes = np.vstack([bboxes_default, bboxes_gt])
neg_marks = np.hstack([marks_default, marks_gt])
classes = np.tile(classes, 2)
scores = np.hstack([scores, scores_gt])
title = "Default boxes: Layer {}".format(
g_ssd_model.feat_layers[i])
visualization.plt_bboxes(img,
classes,
scores,
bboxes,
neg_marks=neg_marks,
title=title)
found_matched = True
return found_matched
image_name = 'Record_01147.bmp' image_name = 'Record_01147_90.jpg' # image_name = 'Record_01639.bmp' # image_name = 'Record_01639_180.jpg' image_name = 'Record_00690_180.jpg' image_name = 'Record_00001.bmp' image_name = 'scream.jpg' image_path = path + image_name # image_name = '../../data/voc/2007_test/VOCdevkit/VOC2007/JPEGImages/000593.jpg' #two cars # image_name = '../../data/voc/2007_test/VOCdevkit/VOC2007/JPEGImages/001086.jpg' #Two person and one bottle # image_name = '../../data/voc/2007_test/VOCdevkit/VOC2007/JPEGImages/009075.jpg'#many people, not very good # image_name = '../../data/voc/2007_test/VOCdevkit/VOC2007/JPEGImages/009957.jpg' #horse and person # image_name = '../../data/voc/2007_test/VOCdevkit/VOC2007/JPEGImages/005575.jpg' #plane # image_name = '../../data/voc/2007_test/VOCdevkit/VOC2007/JPEGImages/004865.jpg' #horse # image_name = '../../data/voc/2007_test/VOCdevkit/VOC2007/JPEGImages/003552.jpg' #car # image_name = '../../data/voc/2007_test/VOCdevkit/VOC2007/JPEGImages/003431.jpg' #one person and one bottle # image_name = '../../data/voc/2007_test/VOCdevkit/VOC2007/JPEGImages/002808.jpg' #motor and car # image_name = '../../data/voc/2007_test/VOCdevkit/VOC2007/JPEGImages/001672.jpg' #train # image_name = '../../data/voc/2007_test/VOCdevkit/VOC2007/JPEGImages/001195.jpg' #dog # image_name = '../../data/voc/2007_test/VOCdevkit/VOC2007/JPEGImages/000568.jpg' #eagle img = mpimg.imread(image_path) rclasses, rscores, rbboxes = process_image(img) # visualization.bboxes_draw_on_img(img, rclasses, rscores, rbboxes, visualization.colors_plasma) visualization.plt_bboxes(img, rclasses, rscores, rbboxes) plt.show()
rclasses, rscores, rbboxes, nms_threshold=nms_threshold)
# Resize bboxes to original image shape. Note: useless for Resize.WARP!
rbboxes = np_methods.bboxes_resize(rbbox_img, rbboxes)
return rclasses, rscores, rbboxes
# Test on some demo image and visualize output.
path = '/home/yjin/data/demo/'
while True:
image_name = input('Enter image name: ')
image_path = Path(path + image_name)
if image_path.is_file():
img = mpimg.imread(image_path)
rclasses, rscores, rbboxes = process_image(img)
# visualization.bboxes_draw_on_img(img, rclasses, rscores, rbboxes, visualization.colors_plasma)
visualization.plt_bboxes(img, rclasses, [1., 1.], rbboxes)
plt.show()
elif image_name == 'exit':
break
else:
print('File does not exist.')
"""
files = os.listdir(path)
for image_name in files:
image_path = Path(path + image_name)
img = mpimg.imread(image_path)
rclasses, rscores, rbboxes = process_image(img)
# visualization.bboxes_draw_on_img(img, rclasses, rscores, rbboxes, visualization.colors_plasma)
visualization.plt_bboxes(img, rclasses, rscores, rbboxes)
plt.show()
"""
