def test_single_gt_single_match_zero_neutral():
anchor_gen = AnchorGenerator(
image_size=(300, 300),
scales=[0.5],
aspect_ratios=[1.0],
clip_boxes=False,
normalize_coordinates=True,
)
anchors = anchor_gen((2, 2))
ground_truth_boxes = tf.constant([[0.24, 0.5, 0.74, 1.0]])
ground_truth_labels = tf.constant([[8]])
matched_gt_boxes, matched_gt_labels, positive_mask, negative_mask = target_assign_func(
ground_truth_boxes,
ground_truth_labels,
anchors,
negative_iou_threshold=1 / 3)
expected_matched_gt_boxes = np.asarray([
anchors[0, :], ground_truth_boxes[0, :], anchors[2, :], anchors[3, :]
])
np.testing.assert_allclose(expected_matched_gt_boxes, matched_gt_boxes)
expected_matched_gt_labels = np.zeros((4, 1))
expected_matched_gt_labels[1] = ground_truth_labels[0]
np.testing.assert_allclose(expected_matched_gt_labels, matched_gt_labels)
expected_positive_mask = np.asarray([0, 1, 0, 0]).astype(np.int)
expected_negative_mask = np.asarray([1, 0, 1, 1]).astype(np.int)
np.testing.assert_equal(expected_positive_mask, positive_mask)
np.testing.assert_equal(expected_negative_mask, negative_mask)
def test_single_gt_no_intersect():
anchor_gen = AnchorGenerator(
image_size=(300, 300),
scales=[0.2],
aspect_ratios=[1.0],
clip_boxes=False,
normalize_coordinates=True,
)
anchors = anchor_gen((2, 2))
ground_truth_boxes = tf.constant([[0.4, 0.65, 0.6, 0.85]])
ground_truth_labels = tf.constant([[8]])
# Since it does not intersect with any anchor, it will be matched with the first gt.
matched_gt_boxes, matched_gt_labels, positive_mask, negative_mask = target_assign_func(
ground_truth_boxes, ground_truth_labels, anchors)
expected_matched_gt_boxes = np.asarray([
ground_truth_boxes[0, :], anchors[1, :], anchors[2, :], anchors[3, :]
])
np.testing.assert_allclose(expected_matched_gt_boxes, matched_gt_boxes)
expected_matched_gt_labels = np.zeros((4, 1))
expected_matched_gt_labels[0] = ground_truth_labels[0]
np.testing.assert_allclose(expected_matched_gt_labels, matched_gt_labels)
expected_positive_mask = np.asarray([1, 0, 0, 0]).astype(np.int)
expected_negative_mask = np.asarray([0, 1, 1, 1]).astype(np.int)
np.testing.assert_equal(expected_positive_mask, positive_mask)
np.testing.assert_equal(expected_negative_mask, negative_mask)
def test_two_gt_two_matches():
anchor_gen = AnchorGenerator(
image_size=(300, 300),
scales=[0.2],
aspect_ratios=[1.0],
clip_boxes=False,
normalize_coordinates=True,
)
anchors = anchor_gen((2, 2))
# The first box will be matched to the second anchor
# The second box will be matched to the first anchor
ground_truth_boxes = tf.constant([
[0.15, 0.65, 0.35, 0.85],
[0.14, 0.64, 0.34, 0.84],
])
ground_truth_labels = tf.constant([[8], [6]])
matched_gt_boxes, matched_gt_labels, positive_mask, negative_mask = target_assign_func(
ground_truth_boxes, ground_truth_labels, anchors)
expected_matched_gt_boxes = np.asarray([
ground_truth_boxes[1, :], ground_truth_boxes[0, :], anchors[2, :],
anchors[3, :]
])
np.testing.assert_allclose(expected_matched_gt_boxes, matched_gt_boxes)
expected_matched_gt_labels = np.zeros((4, 1))
expected_matched_gt_labels[1] = ground_truth_labels[0]
expected_matched_gt_labels[0] = ground_truth_labels[1]
np.testing.assert_allclose(expected_matched_gt_labels, matched_gt_labels)
expected_positive_mask = np.asarray([1, 1, 0, 0]).astype(np.int)
expected_negative_mask = np.asarray([0, 0, 1, 1]).astype(np.int)
np.testing.assert_equal(expected_positive_mask, positive_mask)
np.testing.assert_equal(expected_negative_mask, negative_mask)
def test_single_gt_four_match():
anchor_gen = AnchorGenerator(
image_size=(300, 300),
scales=[0.5],
aspect_ratios=[1.0],
clip_boxes=False,
normalize_coordinates=True,
)
anchors = anchor_gen((2, 2))
ground_truth_boxes = tf.constant([[0.25, 0.25, 0.75, 0.75]])
ground_truth_labels = tf.constant([[8]])
matched_gt_boxes, matched_gt_labels, positive_mask, negative_mask = target_assign_func(
ground_truth_boxes,
ground_truth_labels,
anchors,
positive_iou_threshold=1 / 7,
negative_iou_threshold=1 / 8,
)
expected_matched_gt_boxes = np.tile(ground_truth_boxes, (4, 1))
np.testing.assert_allclose(expected_matched_gt_boxes, matched_gt_boxes)
expected_matched_gt_labels = np.tile(ground_truth_labels, (4, 1))
np.testing.assert_allclose(expected_matched_gt_labels, matched_gt_labels)
expected_positive_mask = np.asarray([1, 1, 1, 1]).astype(np.int)
expected_negative_mask = np.asarray([0, 0, 0, 0]).astype(np.int)
np.testing.assert_equal(expected_positive_mask, positive_mask)
np.testing.assert_equal(expected_negative_mask, negative_mask)
def test_single_gt_best_match():
anchor_gen = AnchorGenerator(
image_size=(300, 300),
scales=[0.2],
aspect_ratios=[1.0],
clip_boxes=False,
normalize_coordinates=True,
)
anchors = anchor_gen((2, 2))
ground_truth_boxes = tf.constant([[0.14, 0.64, 0.34, 0.84]])
ground_truth_labels = tf.constant([[2]], dtype=tf.int64)
matched_gt_boxes, matched_gt_labels, positive_mask, negative_mask = target_assign_tf_func(
ground_truth_boxes, ground_truth_labels, anchors)
matched_gt_boxes = tf.expand_dims(matched_gt_boxes, axis=0)
matched_gt_labels = tf.expand_dims(tf.squeeze(matched_gt_labels, axis=-1),
axis=0)
anchors = tf.expand_dims(anchors, axis=0)
positive_mask = tf.expand_dims(positive_mask, axis=0)
negative_mask = tf.expand_dims(negative_mask, axis=0)
loss_layer = SSDLossLayer()
anchor_labels = tf.constant(
[[[.2, .8, 0.], [.2, .2, .2], [.1, .1, .1], [.1, .1, .1]]],
dtype=tf.float32)
_ = loss_layer(matched_gt_boxes, anchors, matched_gt_labels, anchor_labels,
positive_mask, negative_mask)
reg_losses, cls_losses = loss_layer.losses
# reg loss is 0.5 * (0.01^2+0.01^2+0.01^2+0.01^2) = 0.0002
expected_reg_losses = 0.0002
np.testing.assert_allclose(expected_reg_losses, reg_losses, atol=1e-7)
# cls loss is log(0.2) + log(1/3) + log(1/3) + log(1/3)
expected_cls_losses = 4.905275
np.testing.assert_allclose(expected_cls_losses, cls_losses)
def test_two_gt_two_matches():
anchor_gen = AnchorGenerator(
image_size=(300, 300),
scales=[0.2],
aspect_ratios=[1.0],
clip_boxes=False,
normalize_coordinates=True,
)
anchors = anchor_gen((2, 2))
print(anchors)
# The first box will be matched to the second anchor
# The second box will be matched to the first anchor
ground_truth_boxes = tf.constant([
[0.15, 0.65, 0.35, 0.85],
[0.14, 0.64, 0.34, 0.84],
])
ground_truth_labels = tf.constant([[1], [2]], dtype=tf.int64)
matched_gt_boxes, matched_gt_labels, positive_mask, negative_mask = target_assign_tf_func(
ground_truth_boxes, ground_truth_labels, anchors)
loss_layer = SSDLossLayer(from_logits=True)
anchor_labels = tf.constant(
[[[.2, .8, .0], [.2, .2, .2], [.1, .1, .1], [.1, .1, .1]]],
dtype=tf.float32)
matched_gt_boxes = tf.expand_dims(matched_gt_boxes, axis=0)
matched_gt_labels = tf.expand_dims(tf.squeeze(matched_gt_labels, axis=-1),
axis=0)
anchors = tf.expand_dims(anchors, axis=0)
positive_mask = tf.expand_dims(positive_mask, axis=0)
negative_mask = tf.expand_dims(negative_mask, axis=0)
_ = loss_layer(matched_gt_boxes, anchors, matched_gt_labels, anchor_labels,
positive_mask, negative_mask)
reg_losses, cls_losses = loss_layer.losses
# reg loss is 0.5 * (0.01^2+0.49^2+0.01^2+0.49^2) / 2, since n_positives = 2
np.testing.assert_allclose(0.12009999, reg_losses)
# cls loss is (log(0.22487354) + log(1/3) + log(1/3) + log(1/3)) / 2, since n_positives = 2
np.testing.assert_allclose(2.394027, cls_losses)