def testEncodeDecode(self):
# 1 vs 1 already equal encode and decode.
proposal = generate_gt_boxes(1, image_size=100)
gt_boxes = proposal
deltas = self._encode(proposal, gt_boxes)
decoded_gt_boxes = self._decode(proposal, deltas)
self.assertAllEqual(deltas, np.zeros((1, 4)))
self.assertAllClose(gt_boxes, decoded_gt_boxes)
# 3 vs 3 already equal encode and decode
proposal = generate_gt_boxes(3, image_size=100)
gt_boxes = proposal
deltas = self._encode(proposal, gt_boxes)
decoded_gt_boxes = self._decode(proposal, deltas)
self.assertAllEqual(deltas, np.zeros((3, 4)))
self.assertAllClose(gt_boxes, decoded_gt_boxes)
# 3 vs 4 different encode and decode
proposal = generate_gt_boxes(3, image_size=100)
gt_boxes = generate_gt_boxes(3, image_size=100)
deltas = self._encode(proposal, gt_boxes)
decoded_gt_boxes = self._decode(proposal, deltas)
self.assertAllEqual(deltas.shape, (3, 4))
self.assertAllClose(gt_boxes, decoded_gt_boxes)
def testNonZeroForegrounds(self):
"""Tests we never get zero foregrounds.
We're doing iterations with random gt_boxes and proposals under
conditions that make it likely we would get zero foregrounds if there
is a bug in the code. (Very few gt_boxes and a small minibatch_size).
"""
number_of_iterations = 50
for _ in range(number_of_iterations):
im_shape = np.random.randint(low=600,
high=980,
size=2,
dtype=np.int32)
total_boxes = np.random.randint(low=1, high=4, dtype=np.int32)
total_proposals = np.random.randint(low=4, high=8, dtype=np.int32)
# Generate gt_boxes and then add a label.
gt_boxes = generate_gt_boxes(total_boxes, im_shape)
gt_boxes_w_label = np.concatenate(
[gt_boxes, [[self._placeholder_label]] * total_boxes],
axis=1).astype(np.float32)
# Generate the proposals and add the batch number.
proposed_boxes = generate_gt_boxes(total_proposals, im_shape)
# Run RCNNTarget.
(proposals_label,
_) = self._run_rcnn_target(self._shared_model, gt_boxes_w_label,
proposed_boxes.astype(np.float32))
# Assertion
foreground_number = proposals_label[proposals_label > 0].shape[0]
self.assertGreater(foreground_number, 0)
def _get_image_with_boxes(self, image_size, total_boxes):
image = self._gen_image(*image_size)
bboxes = generate_gt_boxes(
total_boxes,
image_size[:2],
)
return image, bboxes
def testNonZeroForegrounds(self):
"""Tests we never get zero foregrounds.
We're doing iterations with random gt_boxes and proposals under
conditions that make it likely we would get zero foregrounds if there
is a bug in the code. (Very few gt_boxes and a small minibatch_size).
"""
number_of_iterations = 50
for _ in range(number_of_iterations):
im_shape = np.random.randint(
low=600, high=980, size=2, dtype=np.int32
)
total_boxes = np.random.randint(
low=1, high=4, dtype=np.int32
)
total_proposals = np.random.randint(
low=4, high=8, dtype=np.int32
)
# Generate gt_boxes and then add a label.
gt_boxes = generate_gt_boxes(
total_boxes, im_shape
)
gt_boxes_w_label = np.concatenate(
[gt_boxes, [[self._placeholder_label]] * total_boxes],
axis=1
).astype(np.float32)
# Generate the proposals and add the batch number.
proposed_boxes = generate_gt_boxes(
total_proposals, im_shape
)
# Run RCNNTarget.
(proposals_label, _) = self._run_rcnn_target(
self._shared_model, gt_boxes_w_label,
proposed_boxes.astype(np.float32)
)
# Assertion
foreground_number = proposals_label[proposals_label > 0].shape[0]
self.assertGreater(foreground_number, 0)
def testEncodeDecodeRandomizedValues(self):
for i in range(1, 2000, 117):
gt_boxes = generate_gt_boxes(i, image_size=800)
proposals = generate_gt_boxes(i, image_size=800)
self._encode_decode(proposals, gt_boxes)
def _get_image_with_boxes(self, image_size, total_boxes):
image = self._gen_image(*image_size)
bboxes = generate_gt_boxes(
total_boxes, image_size[:2],
)
return image, bboxes
