octave_base_scale=6,
scales_per_octave=1,
)
featmap_sizes = [
[80, 80],
[40, 40],
[20, 20],
[10, 10],
[5, 5],
]
mlvl_anchors = anchor_gen.grid_anchors(featmap_sizes)
num_level_bboxes = [a.shape[0] for a in mlvl_anchors]
anchors = tf.concat(mlvl_anchors, axis=0)
bbox_coder = BBoxCoder(anchors)
@curry
def preprocess(example,
output_size=(HEIGHT, WIDTH),
max_objects=50,
training=True):
image, objects, image_id = decode(example)
# if training:
# image = random_resize(image, output_size, ratio_range=(0.8, 1.2))
# image, objects = random_crop(image, objects, output_size)
# image, objects = random_hflip(image, objects, 0.5)
# else:
image = resize(image, output_size)
anchor_gen = AnchorGenerator(
strides=[8, 16, 32, 64, 128],
ratios=[0.5, 1.0, 2.0],
octave_base_scale=2,
scales_per_octave=3,
)
featmap_sizes = [
[32, 32], [16, 16], [8, 8], [4, 4], [2, 2],
]
mlvl_anchors = anchor_gen.grid_anchors(featmap_sizes)
num_level_bboxes = [a.shape[0] for a in mlvl_anchors]
anchors = tf.concat(mlvl_anchors, axis=0)
bbox_coder = BBoxCoder(anchors, (0.1, 0.1, 0.2, 0.2))
@curry
def preprocess(example, output_size=(HEIGHT, WIDTH), max_objects=50, training=True):
image, objects, image_id = decode(example)
if training:
image = random_resize(image, output_size, ratio_range=(0.5, 2.0))
image, objects = random_crop(image, objects, output_size)
image, objects = random_hflip(image, objects, 0.5)
else:
image = resize(image, output_size)
image = normalize(image, [123.68, 116.779, 103.939], [58.393, 57.12, 57.375])
image, objects = pad_to(image, objects, output_size)