Skip to content

Commit

Permalink
Update tests for new vectorized distance functions.
Browse files Browse the repository at this point in the history
  • Loading branch information
@facundo-lezama
facundo-lezama committed Nov 2, 2022
1 parent 054da90 commit 703bf5d
Show file tree
Hide file tree
Showing 4 changed files with 109 additions and 54 deletions.
6 changes: 4 additions & 2 deletions tests/conftest.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -7,7 +7,7 @@
@pytest.fixture
def mock_det():
class FakeDetection:
def __init__(self, points, scores=None) -> None:
def __init__(self, points, scores=None, label=None) -> None:
if not isinstance(points, np.ndarray):
points = np.array(points)
self.points = points
Expand All @@ -17,19 +17,21 @@ def __init__(self, points, scores=None) -> None:
if scores.ndim == 0 and points.shape[0] > 1:
scores = np.full(points.shape[0], scores)
self.scores = scores
self.label = label

return FakeDetection


@pytest.fixture
def mock_obj(mock_det):
class FakeTrackedObject:
def __init__(self, points, scores=None):
def __init__(self, points, scores=None, label=None):
if not isinstance(points, np.ndarray):
points = np.array(points)

self.estimate = points
self.last_detection = mock_det(points, scores=scores)
self.label = label

return FakeTrackedObject

Expand Down
3 changes: 1 addition & 2 deletions tests/mot_metrics.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -4,7 +4,6 @@
import pandas as pd

from norfair import FilterPyKalmanFilterFactory, Tracker, metrics
from norfair.distances import iou

DATASET_PATH = "train"
MOTA_ERROR_THRESHOLD = 0.0
Expand Down Expand Up @@ -48,7 +47,7 @@ def mot_metrics():
)

tracker = Tracker(
distance_function=iou,
distance_function="iou",
distance_threshold=DISTANCE_THRESHOLD,
detection_threshold=DETECTION_THRESHOLD,
pointwise_hit_counter_max=POINTWISE_HIT_COUNTER_MAX,
Expand Down
132 changes: 93 additions & 39 deletions tests/test_distances.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -2,8 +2,12 @@
import pytest

from norfair.distances import (
ScalarDistance,
ScipyDistance,
VectorizedDistance,
create_keypoints_voting_distance,
create_normalized_mean_euclidean_distance,
frobenius,
get_distance_by_name,
)

Expand All @@ -14,37 +18,37 @@ def test_frobenius(mock_obj, mock_det):
# perfect match
det = mock_det([[1, 2], [3, 4]])
obj = mock_obj([[1, 2], [3, 4]])
np.testing.assert_almost_equal(fro(det, obj), 0)
np.testing.assert_almost_equal(fro.distance_function(det, obj), 0)

# foat type
det = mock_det([[1.1, 2.2], [3.3, 4.4]])
obj = mock_obj([[1.1, 2.2], [3.3, 4.4]])
np.testing.assert_almost_equal(fro(det, obj), 0)
np.testing.assert_almost_equal(fro.distance_function(det, obj), 0)

# distance of 1 in 1 dimension of 1 point
det = mock_det([[1, 2], [3, 4]])
obj = mock_obj([[2, 2], [3, 4]])
np.testing.assert_almost_equal(fro(det, obj), np.sqrt(1))
np.testing.assert_almost_equal(fro.distance_function(det, obj), np.sqrt(1))

# distance of 2 in 1 dimension of 1 point
det = mock_det([[1, 2], [3, 4]])
obj = mock_obj([[3, 2], [3, 4]])
np.testing.assert_almost_equal(fro(det, obj), 2)
np.testing.assert_almost_equal(fro.distance_function(det, obj), 2)

# distance of 1 in all dimensions of all points
det = mock_det([[1, 2], [3, 4]])
obj = mock_obj([[2, 3], [4, 5]])
np.testing.assert_almost_equal(fro(det, obj), np.sqrt(4))
np.testing.assert_almost_equal(fro.distance_function(det, obj), np.sqrt(4))

# negative difference
det = mock_det([[1, 2], [3, 4]])
obj = mock_obj([[-1, 2], [3, 4]])
np.testing.assert_almost_equal(fro(det, obj), 2)
np.testing.assert_almost_equal(fro.distance_function(det, obj), 2)

# negative equals
det = mock_det([[-1, 2], [3, 4]])
obj = mock_obj([[-1, 2], [3, 4]])
np.testing.assert_almost_equal(fro(det, obj), 0)
np.testing.assert_almost_equal(fro.distance_function(det, obj), 0)


def test_mean_manhattan(mock_det, mock_obj):
Expand All @@ -53,37 +57,37 @@ def test_mean_manhattan(mock_det, mock_obj):
# perfect match
det = mock_det([[1, 2], [3, 4]])
obj = mock_obj([[1, 2], [3, 4]])
np.testing.assert_almost_equal(man(det, obj), 0)
np.testing.assert_almost_equal(man.distance_function(det, obj), 0)

# foat type
det = mock_det([[1.1, 2.2], [3.3, 4.4]])
obj = mock_obj([[1.1, 2.2], [3.3, 4.4]])
np.testing.assert_almost_equal(man(det, obj), 0)
np.testing.assert_almost_equal(man.distance_function(det, obj), 0)

# distance of 1 in 1 dimension of 1 point
det = mock_det([[1, 2], [3, 4]])
obj = mock_obj([[2, 2], [3, 4]])
np.testing.assert_almost_equal(man(det, obj), 1 / 2)
np.testing.assert_almost_equal(man.distance_function(det, obj), 1 / 2)

# distance of 2 in 1 dimension of 1 point
det = mock_det([[1, 2], [3, 4]])
obj = mock_obj([[3, 2], [3, 4]])
np.testing.assert_almost_equal(man(det, obj), 1)
np.testing.assert_almost_equal(man.distance_function(det, obj), 1)

# distance of 1 in all dimensions of all points
det = mock_det([[1, 2], [3, 4]])
obj = mock_obj([[2, 3], [4, 5]])
np.testing.assert_almost_equal(man(det, obj), 2)
np.testing.assert_almost_equal(man.distance_function(det, obj), 2)

# negative difference
det = mock_det([[1, 2], [3, 4]])
obj = mock_obj([[-1, 2], [3, 4]])
np.testing.assert_almost_equal(man(det, obj), 1)
np.testing.assert_almost_equal(man.distance_function(det, obj), 1)

# negative equals
det = mock_det([[-1, 2], [3, 4]])
obj = mock_obj([[-1, 2], [3, 4]])
np.testing.assert_almost_equal(man(det, obj), 0)
np.testing.assert_almost_equal(man.distance_function(det, obj), 0)


def test_mean_euclidean(mock_det, mock_obj):
Expand All @@ -92,47 +96,47 @@ def test_mean_euclidean(mock_det, mock_obj):
# perfect match
det = mock_det([[1, 2], [3, 4]])
obj = mock_obj([[1, 2], [3, 4]])
np.testing.assert_almost_equal(euc(det, obj), 0)
np.testing.assert_almost_equal(euc.distance_function(det, obj), 0)

# foat type
det = mock_det([[1.1, 2.2], [3.3, 4.4]])
obj = mock_obj([[1.1, 2.2], [3.3, 4.4]])
np.testing.assert_almost_equal(euc(det, obj), 0)
np.testing.assert_almost_equal(euc.distance_function(det, obj), 0)

# distance of 1 in 1 dimension of 1 point
det = mock_det([[1, 2], [3, 4]])
obj = mock_obj([[2, 2], [3, 4]])
np.testing.assert_almost_equal(euc(det, obj), 1 / 2)
np.testing.assert_almost_equal(euc.distance_function(det, obj), 1 / 2)

# distance of 2 in 1 dimension of 1 point
det = mock_det([[1, 2], [3, 4]])
obj = mock_obj([[3, 2], [3, 4]])
np.testing.assert_almost_equal(euc(det, obj), 1)
np.testing.assert_almost_equal(euc.distance_function(det, obj), 1)

# distance of 2 in 1 dimension of all points
det = mock_det([[1, 2], [3, 4]])
obj = mock_obj([[3, 2], [5, 4]])
np.testing.assert_almost_equal(euc(det, obj), 2)
np.testing.assert_almost_equal(euc.distance_function(det, obj), 2)

# distance of 2 in all dimensions of all points
det = mock_det([[1, 2], [3, 4]])
obj = mock_obj([[3, 4], [5, 6]])
np.testing.assert_almost_equal(euc(det, obj), np.sqrt(8))
np.testing.assert_almost_equal(euc.distance_function(det, obj), np.sqrt(8))

# distance of 1 in all dimensions of all points
det = mock_det([[1, 2], [3, 4]])
obj = mock_obj([[2, 3], [4, 5]])
np.testing.assert_almost_equal(euc(det, obj), np.sqrt(2))
np.testing.assert_almost_equal(euc.distance_function(det, obj), np.sqrt(2))

# negative difference
det = mock_det([[1, 2], [3, 4]])
obj = mock_obj([[-1, 2], [3, 4]])
np.testing.assert_almost_equal(euc(det, obj), 1)
np.testing.assert_almost_equal(euc.distance_function(det, obj), 1)

# negative equals
det = mock_det([[-1, 2], [3, 4]])
obj = mock_obj([[-1, 2], [3, 4]])
np.testing.assert_almost_equal(euc(det, obj), 0)
np.testing.assert_almost_equal(euc.distance_function(det, obj), 0)


def test_iou(mock_det, mock_obj):
Expand All @@ -142,60 +146,60 @@ def test_iou(mock_det, mock_obj):
# perfect match
det = mock_det([[0, 0], [1, 1]])
obj = mock_obj([[0, 0], [1, 1]])
np.testing.assert_almost_equal(iou(det, obj), 0)
np.testing.assert_almost_equal(iou_opt(det, obj), 0)
np.testing.assert_almost_equal(iou.distance_function(det, obj), 0)
np.testing.assert_almost_equal(iou_opt.distance_function(det, obj), 0)

# float type
det = mock_det([[0.0, 0.0], [1.1, 1.1]])
obj = mock_obj([[0.0, 0.0], [1.1, 1.1]])
np.testing.assert_almost_equal(iou(det, obj), 0)
np.testing.assert_almost_equal(iou_opt(det, obj), 0)
np.testing.assert_almost_equal(iou.distance_function(det, obj), 0)
np.testing.assert_almost_equal(iou_opt.distance_function(det, obj), 0)

# det contained in obj
det = mock_det([[0, 0], [1, 1]])
obj = mock_obj([[0, 0], [2, 2]])
np.testing.assert_almost_equal(iou(det, obj), 1 - 1 / 4)
np.testing.assert_almost_equal(iou_opt(det, obj), 1 - 1 / 4)
np.testing.assert_almost_equal(iou.distance_function(det, obj), 1 - 1 / 4)
np.testing.assert_almost_equal(iou_opt.distance_function(det, obj), 1 - 1 / 4)

# no overlap
det = mock_det([[0, 0], [1, 1]])
obj = mock_obj([[1, 1], [2, 2]])
np.testing.assert_almost_equal(iou(det, obj), 1)
np.testing.assert_almost_equal(iou_opt(det, obj), 1)
np.testing.assert_almost_equal(iou.distance_function(det, obj), 1)
np.testing.assert_almost_equal(iou_opt.distance_function(det, obj), 1)

# obj fully contained on det
det = mock_det([[0, 0], [4, 4]])
obj = mock_obj([[1, 1], [2, 2]])
np.testing.assert_almost_equal(iou(det, obj), 1 - 1 / 16)
np.testing.assert_almost_equal(iou_opt(det, obj), 1 - 1 / 16)
np.testing.assert_almost_equal(iou.distance_function(det, obj), 1 - 1 / 16)
np.testing.assert_almost_equal(iou_opt.distance_function(det, obj), 1 - 1 / 16)

# partial overlap
det = mock_det([[0, 0], [2, 2]])
obj = mock_obj([[1, 1], [3, 3]])
np.testing.assert_almost_equal(iou(det, obj), 1 - 1 / (8 - 1))
np.testing.assert_almost_equal(iou_opt(det, obj), 1 - 1 / (8 - 1))
np.testing.assert_almost_equal(iou.distance_function(det, obj), 1 - 1 / (8 - 1))
np.testing.assert_almost_equal(iou_opt.distance_function(det, obj), 1 - 1 / (8 - 1))

# invalid bbox
det = mock_det([[0, 0]])
obj = mock_obj([[0, 0]])
with pytest.raises(AssertionError):
iou(det, obj)
iou.distance_function(det, obj)

# invalid bbox
det = mock_det([[0, 0], [1, 1], [2, 2]])
obj = mock_obj([[0, 0], [2, 2]])
with pytest.raises(AssertionError):
iou(det, obj)
iou.distance_function(det, obj)

# invalid box should be corrected
det = mock_det([[4, 4], [0, 0]])
obj = mock_obj([[0, 0], [4, 4]])
np.testing.assert_almost_equal(iou(det, obj), 0)
np.testing.assert_almost_equal(iou.distance_function(det, obj), 0)

# invalid box should be corrected
det = mock_det([[0, 0], [4, 4]])
obj = mock_obj([[4, 4], [0, 0]])
np.testing.assert_almost_equal(iou(det, obj), 0)
np.testing.assert_almost_equal(iou.distance_function(det, obj), 0)


def test_keypoint_vote(mock_obj, mock_det):
Expand Down Expand Up @@ -286,3 +290,53 @@ def test_normalized_euclidean(mock_obj, mock_det):
det = mock_det([[-1, 2], [3, 4]])
obj = mock_obj([[-1, 2], [3, 4]])
np.testing.assert_almost_equal(norm_e(det, obj), 0)


def test_scalar_distance(mock_obj, mock_det):
fro = ScalarDistance(frobenius)

det = mock_det([[1, 2], [3, 4]])
obj = mock_obj([[1, 2], [3, 4]])

dist_matrix = fro.get_distances([obj], [det])

assert type(dist_matrix) == np.ndarray
assert dist_matrix.shape == (1, 1)
assert dist_matrix[0, 0] == 0


def test_vectorized_distance(mock_obj, mock_det):
def distance_function(cands, objs):
distance_matrix = np.full(
(len(cands), len(objs)),
fill_value=np.inf,
dtype=np.float32,
)
for c, cand in enumerate(cands):
for o, obj in enumerate(objs):
distance_matrix[c, o] = np.linalg.norm(cand - obj)
return distance_matrix

fro = VectorizedDistance(distance_function)

det = mock_det([[1, 2], [3, 4]])
obj = mock_obj([[1, 2], [3, 4]])

dist_matrix = fro.get_distances([obj], [det])

assert type(dist_matrix) == np.ndarray
assert dist_matrix.shape == (1, 1)
assert dist_matrix[0, 0] == 0


def test_scipy_distance(mock_obj, mock_det):
euc = ScipyDistance("euclidean")

det = mock_det([[1, 2], [3, 4]])
obj = mock_obj([[1, 2], [4, 4]])

dist_matrix = euc.get_distances([obj], [det])

assert type(dist_matrix) == np.ndarray
assert dist_matrix.shape == (1, 1)
assert dist_matrix[0, 0] == 1.0
Loading

0 comments on commit 703bf5d

Please sign in to comment.