def test_multi_feature_maps_customized_stride():
anchor_gen = multi_scale_anchor_generator.MultiScaleAnchorGenerator(
image_size=(300, 300),
scales=[[0.1], [0.2]],
aspect_ratios=[[1.0], [1.0]],
strides=[[120, 120], [160, 160]],
clip_boxes=False,
normalize_coordinates=False,
)
anchor_out = anchor_gen([(3, 3), (2, 2)])
# The first center of anchor point for the first feature map is 120 * 0.5 = 60, then 180
# The first center of anchor point for the second feature map is 160 * 0.5 = 80, then 240
expected_out = np.asarray(
[
[45.0, 45.0, 75.0, 75.0],
[45.0, 165.0, 75.0, 195.0],
[45.0, 285.0, 75.0, 315.0],
[165.0, 45.0, 195.0, 75.0],
[165.0, 165.0, 195.0, 195.0],
[165.0, 285.0, 195.0, 315.0],
[285.0, 45.0, 315.0, 75.0],
[285.0, 165.0, 315.0, 195.0],
[285.0, 285.0, 315.0, 315.0],
[50.0, 50.0, 110.0, 110.0],
[50.0, 210.0, 110.0, 270.0],
[210.0, 50.0, 270.0, 110.0],
[210.0, 210.0, 270.0, 270.0],
]
).astype(np.float32)
np.testing.assert_allclose(expected_out, anchor_out)
def test_multi_feature_maps_absolute_coordinate():
anchor_gen = multi_scale_anchor_generator.MultiScaleAnchorGenerator(
image_size=(300, 300),
scales=[[0.1], [0.2]],
aspect_ratios=[[1.0], [1.0]],
clip_boxes=False,
normalize_coordinates=False,
)
anchor_out = anchor_gen([(3, 3), (2, 2)])
# The first height and width is 30, the second height and width is 60.
expected_out = np.asarray(
[
[35.0, 35.0, 65.0, 65.0],
[35.0, 135.0, 65.0, 165.0],
[35.0, 235.0, 65.0, 265.0],
[135.0, 35.0, 165.0, 65.0],
[135.0, 135.0, 165.0, 165.0],
[135.0, 235.0, 165.0, 265.0],
[235.0, 35.0, 265.0, 65.0],
[235.0, 135.0, 265.0, 165.0],
[235.0, 235.0, 265.0, 265.0],
[45.0, 45.0, 105.0, 105.0],
[45.0, 195.0, 105.0, 255.0],
[195.0, 45.0, 255.0, 105.0],
[195.0, 195.0, 255.0, 255.0],
]
).astype(np.float32)
np.testing.assert_allclose(expected_out, anchor_out)
def test_config_with_custom_name():
layer = multi_scale_anchor_generator.MultiScaleAnchorGenerator(
(300, 300), [[1.0]], [[1.0]], name="multi_anchor_generator"
)
config = layer.get_config()
layer_1 = multi_scale_anchor_generator.MultiScaleAnchorGenerator.from_config(config)
np.testing.assert_equal(layer_1.name, layer.name)
def test_multi_feature_maps_customized_offset():
anchor_gen = multi_scale_anchor_generator.MultiScaleAnchorGenerator(
image_size=(300, 300),
scales=[[0.1], [0.2]],
aspect_ratios=[[1.0], [1.0]],
offsets=[[0.2, 0.2], [0.3, 0.3]],
clip_boxes=False,
normalize_coordinates=False,
)
anchor_out = anchor_gen([(3, 3), (2, 2)])
# The first center of anchor point for the first feature map is 100 * 0.2 = 20, then 120
# The first center of anchor point for the second feature map is 150 * 0.3 = 45, then 195
expected_out = np.asarray(
[
[5.0, 5.0, 35.0, 35.0],
[5.0, 105.0, 35.0, 135.0],
[5.0, 205.0, 35.0, 235.0],
[105.0, 5.0, 135.0, 35.0],
[105.0, 105.0, 135.0, 135.0],
[105.0, 205.0, 135.0, 235.0],
[205.0, 5.0, 235.0, 35.0],
[205.0, 105.0, 235.0, 135.0],
[205.0, 205.0, 235.0, 235.0],
[15, 15, 75, 75],
[15, 165, 75, 225],
[165, 15, 225, 75],
[165, 165, 225, 225],
]
).astype(np.float32)
np.testing.assert_allclose(expected_out, anchor_out)
def test_single_feature_map_absolute_coordinate():
anchor_gen = multi_scale_anchor_generator.MultiScaleAnchorGenerator(
scales=[[0.2]],
aspect_ratios=[[1.0]],
clip_boxes=False,
normalize_coordinates=False,
)
anchor_out = anchor_gen((300, 300), [(2, 2)])
expected_out = np.asarray([
[45, 45, 105, 105],
[45, 195, 105, 255],
[195, 45, 255, 105],
[195, 195, 255, 255],
]).astype(np.float32)
np.testing.assert_allclose(expected_out, anchor_out)
def test_single_feature_map_multi_aspect_ratios():
anchor_gen = multi_scale_anchor_generator.MultiScaleAnchorGenerator(
scales=[[0.2, 0.2]],
aspect_ratios=[[0.64, 1.0]],
clip_boxes=False,
normalize_coordinates=False,
)
anchor_out = anchor_gen((300, 300), [(2, 2)])
# height is 300 * 0.2 / 0.8 = 75; width is 300 * 0.2 * 0.8 = 48
expected_out = np.asarray([
[37.5, 51.0, 112.5, 99.0],
[45.0, 45.0, 105.0, 105.0],
[37.5, 201.0, 112.5, 249.0],
[45.0, 195.0, 105.0, 255.0],
[187.5, 51.0, 262.5, 99.0],
[195.0, 45.0, 255.0, 105.0],
[187.5, 201.0, 262.5, 249.0],
[195.0, 195.0, 255.0, 255.0],
]).astype(np.float32)
np.testing.assert_allclose(expected_out, anchor_out)
