def __init__(self, num_categories, ignored_label, max_instances_per_category,
offset, is_thing=None, rescale_predictions=False):
"""Constructs Panoptic Quality evaluation class.
The class provides the interface to Panoptic Quality metrics_fn.
Args:
num_categories: The number of segmentation categories (or "classes" in the
dataset.
ignored_label: A category id that is ignored in evaluation, e.g. the void
label as defined in COCO panoptic segmentation dataset.
max_instances_per_category: The maximum number of instances for each
category. Used in ensuring unique instance labels.
offset: The maximum number of unique labels. This is used, by multiplying
the ground-truth labels, to generate unique ids for individual regions
of overlap between groundtruth and predicted segments.
is_thing: A boolean array of length `num_categories`. The entry
`is_thing[category_id]` is True iff that category is a "thing" category
instead of "stuff." Default to `None`, and it means categories are not
classified into these two categories.
rescale_predictions: `bool`, whether to scale back prediction to original
image sizes. If True, groundtruths['image_info'] is used to rescale
predictions.
"""
self._pq_metric_module = panoptic_quality.PanopticQuality(
num_categories, ignored_label, max_instances_per_category, offset)
self._is_thing = is_thing
self._rescale_predictions = rescale_predictions
self._required_prediction_fields = ['category_mask', 'instance_mask']
self._required_groundtruth_fields = ['category_mask', 'instance_mask']
self.reset_states()
def test_instance_order_is_arbitrary(self):
category_mask = np.array([
[1, 1, 1, 1, 1, 1],
[1, 1, 1, 1, 1, 1],
[1, 2, 2, 1, 2, 2],
[1, 2, 2, 1, 2, 2],
[1, 1, 1, 1, 1, 1],
[1, 1, 1, 1, 1, 1],
],
dtype=np.uint16)
groundtruth_instance_mask = np.array([
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 1, 1, 0, 0, 0],
[0, 1, 1, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
],
dtype=np.uint16)
predicted_instance_mask = np.array([
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 1, 1],
[0, 0, 0, 0, 1, 1],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
],
dtype=np.uint16)
groundtruths = {
'category_mask': category_mask,
'instance_mask': groundtruth_instance_mask
}
predictions = {
'category_mask': category_mask,
'instance_mask': predicted_instance_mask
}
pq_metric = panoptic_quality.PanopticQuality(
num_categories=3,
ignored_label=0,
max_instances_per_category=10,
offset=100)
pq_metric.compare_and_accumulate(groundtruths, predictions)
np.testing.assert_array_equal(pq_metric.iou_per_class, [0.0, 1.0, 2.0])
np.testing.assert_array_equal(pq_metric.tp_per_class, [0, 1, 2])
np.testing.assert_array_equal(pq_metric.fn_per_class, [0, 0, 0])
np.testing.assert_array_equal(pq_metric.fp_per_class, [0, 0, 0])
results = pq_metric.result()
np.testing.assert_array_equal(results['pq_per_class'], [0.0, 1.0, 1.0])
np.testing.assert_array_equal(results['rq_per_class'], [0.0, 1.0, 1.0])
np.testing.assert_array_equal(results['sq_per_class'], [0.0, 1.0, 1.0])
self.assertAlmostEqual(results['All_pq'], 1.0)
self.assertAlmostEqual(results['All_rq'], 1.0)
self.assertAlmostEqual(results['All_sq'], 1.0)
self.assertEqual(results['All_num_categories'], 2)
def test_totally_wrong(self):
category_mask = np.array([
[0, 0, 0, 0, 0, 0],
[0, 1, 0, 0, 1, 0],
[0, 1, 1, 1, 1, 0],
[0, 1, 1, 1, 1, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
],
dtype=np.uint16)
instance_mask = np.zeros([6, 6], np.uint16)
groundtruths = {
'category_mask': category_mask,
'instance_mask': instance_mask
}
predictions = {
'category_mask': 1 - category_mask,
'instance_mask': instance_mask
}
pq_metric = panoptic_quality.PanopticQuality(
num_categories=2,
ignored_label=2,
max_instances_per_category=1,
offset=16)
pq_metric.compare_and_accumulate(groundtruths, predictions)
np.testing.assert_array_equal(pq_metric.iou_per_class, [0.0, 0.0])
np.testing.assert_array_equal(pq_metric.tp_per_class, [0, 0])
np.testing.assert_array_equal(pq_metric.fn_per_class, [1, 1])
np.testing.assert_array_equal(pq_metric.fp_per_class, [1, 1])
results = pq_metric.result()
np.testing.assert_array_equal(results['pq_per_class'], [0.0, 0.0])
np.testing.assert_array_equal(results['rq_per_class'], [0.0, 0.0])
np.testing.assert_array_equal(results['sq_per_class'], [0.0, 0.0])
self.assertAlmostEqual(results['All_pq'], 0.0)
self.assertAlmostEqual(results['All_rq'], 0.0)
self.assertAlmostEqual(results['All_sq'], 0.0)
self.assertEqual(results['All_num_categories'], 2)
def test_matches_by_iou(self):
groundtruth_instance_mask = np.array([
[1, 1, 1, 1, 1, 1],
[1, 1, 1, 1, 1, 1],
[1, 1, 2, 2, 2, 1],
[1, 2, 2, 2, 2, 1],
[1, 1, 1, 1, 1, 1],
[1, 1, 1, 1, 1, 1],
],
dtype=np.uint16)
good_det_instance_mask = np.array([
[1, 1, 1, 1, 1, 1],
[1, 1, 1, 1, 1, 1],
[1, 2, 2, 2, 2, 1],
[1, 2, 2, 2, 1, 1],
[1, 1, 1, 1, 1, 1],
[1, 1, 1, 1, 1, 1],
],
dtype=np.uint16)
groundtruths = {
'category_mask': np.zeros_like(groundtruth_instance_mask),
'instance_mask': groundtruth_instance_mask
}
predictions = {
'category_mask': np.zeros_like(good_det_instance_mask),
'instance_mask': good_det_instance_mask
}
pq_metric = panoptic_quality.PanopticQuality(
num_categories=1,
ignored_label=2,
max_instances_per_category=16,
offset=16)
pq_metric.compare_and_accumulate(groundtruths, predictions)
# iou(1, 1) = 28/30
# iou(2, 2) = 6 / 8
np.testing.assert_array_almost_equal(pq_metric.iou_per_class,
[28 / 30 + 6 / 8])
np.testing.assert_array_equal(pq_metric.tp_per_class, [2])
np.testing.assert_array_equal(pq_metric.fn_per_class, [0])
np.testing.assert_array_equal(pq_metric.fp_per_class, [0])
results = pq_metric.result()
np.testing.assert_array_equal(results['pq_per_class'],
[(28 / 30 + 6 / 8) / 2])
np.testing.assert_array_equal(results['rq_per_class'], [1.0])
np.testing.assert_array_equal(results['sq_per_class'],
[(28 / 30 + 6 / 8) / 2])
self.assertAlmostEqual(results['All_pq'], (28 / 30 + 6 / 8) / 2)
self.assertAlmostEqual(results['All_rq'], 1.0)
self.assertAlmostEqual(results['All_sq'], (28 / 30 + 6 / 8) / 2)
self.assertEqual(results['All_num_categories'], 1)
bad_det_instance_mask = np.array([
[1, 1, 1, 1, 1, 1],
[1, 1, 1, 1, 1, 1],
[1, 1, 1, 2, 2, 1],
[1, 1, 1, 2, 2, 1],
[1, 1, 1, 2, 2, 1],
[1, 1, 1, 1, 1, 1],
],
dtype=np.uint16)
predictions['instance_mask'] = bad_det_instance_mask
pq_metric.reset()
pq_metric.compare_and_accumulate(groundtruths, predictions)
# iou(1, 1) = 27/32
np.testing.assert_array_almost_equal(pq_metric.iou_per_class,
[27 / 32])
np.testing.assert_array_equal(pq_metric.tp_per_class, [1])
np.testing.assert_array_equal(pq_metric.fn_per_class, [1])
np.testing.assert_array_equal(pq_metric.fp_per_class, [1])
results = pq_metric.result()
np.testing.assert_array_equal(results['pq_per_class'], [27 / 32 / 2])
np.testing.assert_array_equal(results['rq_per_class'], [0.5])
np.testing.assert_array_equal(results['sq_per_class'], [27 / 32])
self.assertAlmostEqual(results['All_pq'], 27 / 32 / 2)
self.assertAlmostEqual(results['All_rq'], 0.5)
self.assertAlmostEqual(results['All_sq'], 27 / 32)
self.assertEqual(results['All_num_categories'], 1)
