)
def parse_example(example):
return get_resized_image(example), get_ground_truth_boxes(example)
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Visualizes the contents of TFRecord files in TensorBoard')
parser.add_argument('wildcard', help='The wildcard to find TFRecord files to read')
parser.add_argument('--out', type=str, default='logs', help='The output directory for summaries')
parser.add_argument('--batch', type=int, default=100, help='The batch size of each image batch to display')
parser.add_argument('--max-imgs', type=int, default=3, help='The numbers of images to display in each batch')
args = parser.parse_args()
AUTO = tf.data.experimental.AUTOTUNE
filenames = tf.data.Dataset.list_files(args.wildcard)
example_protos = tf.data.TFRecordDataset(filenames, num_parallel_reads=AUTO)
examples = example_protos.map(parse_example_proto, num_parallel_calls=AUTO)
data = examples.map(parse_example, num_parallel_calls=AUTO)
data = data.padded_batch(args.batch, padded_shapes=((None, None, None), (None, None)))
writer = tf.summary.create_file_writer(args.out)
i = 0
for image_batch, gt_batch in data:
print(f'[+] Processing batch {i}')
image_batch = draw_bounding_boxes(image_batch, convert_bounding_boxes_to_tf_format(gt_batch))
with writer.as_default():
draw_image_batch(f'Batch {i}', image_batch, max_images=args.max_imgs)
i += 1
writer.flush()
default=0,
help='ID of the image to visualize')
parser.add_argument('--out',
type=str,
default='logs',
help='The output directory for summaries')
args = parser.parse_args()
# load the kaggle annotations
kaggle_annotations = load_kaggle_annotations(args.labels, args.boxes)
scut_annotations = load_scut_annotations(args.scut_annotations_folder)
annotations = aggregate_annotations(
[kaggle_annotations, scut_annotations],
[args.kaggle_basepath, args.scut_basepath])
# process the first image
image_annotation = annotations[args.id]
image = tf.image.decode_jpeg(
tf.io.read_file(image_annotation["image_filename"]))
image = tf.image.convert_image_dtype(image, tf.float32)
image = tf.image.resize(image, (480, 640))
image_batch = tf.expand_dims(image, axis=0)
writer = tf.summary.create_file_writer(args.out)
with writer.as_default():
draw_image_batch('Image', image_batch)
writer.flush()
model.load_weights(args.model_path)
labels, deltas = model.predict(image_batch)
bboxes = get_bounding_boxes_from_labels(labels, config)
bboxes = apply_deltas_to_bounding_boxes(bboxes,
tf.reshape(deltas,
[-1, 4]), config)
bboxes = convert_bounding_boxes_to_tf_format(bboxes)
bboxes = tf.squeeze(
bboxes,
[0]) # apparently after all the efforts done working with bbox batches
# tensorflow betrayed us and decided that nms should work on
# non batched boxes. lame
selected_indices = tf.image.non_max_suppression(bboxes,
tf.reshape(labels, [-1]),
tf.shape(bboxes)[0],
iou_threshold=0.3)
selected_boxes = tf.gather(bboxes, selected_indices)
selected_boxes = tf.expand_dims(
selected_boxes, axis=0) # now tensorflow wants the batches again
# so lame
output_image = draw_bounding_boxes(image_batch, selected_boxes)
writer = tf.summary.create_file_writer(args.out)
with writer.as_default():
draw_image_batch('Predicted boxes', output_image)
writer.flush()