def test_ious():
b1 = [[10, 20]]
b2 = [[10, 21], [30, 40]]
iou_1d = calculate_iou(b1, b2, dim=1)
assert_almost_equal(iou_1d, np.array([[0.9091, 0.]]))
b1 = [[20.1, 20.1, 30.1, 30.1], [15, 15, 25, 25]]
b2 = [[10, 10, 20, 20]]
iou_2d = calculate_iou(b1, b2, dim=2)
assert_almost_equal(iou_2d, [[0], [0.1429]])
b1 = [[10, 10, 10, 20, 20, 20]]
b2 = [[10, 11, 10.2, 21, 19.9, 20.3], [30, 30, 30, 90, 90, 90]]
iou_3d = calculate_iou(b1, b2, dim=3)
assert_almost_equal(iou_3d, [[0.7811, 0]])
def cost_matrix_iou_feature(trackers: Sequence[Tracker],
detections: Sequence[Detection],
feature_similarity_fn=angular_similarity,
feature_similarity_beta: float = None):
# boxes
b1 = np.array([t.box for t in trackers])
b2 = np.array([d.box for d in detections])
# box iou
inferred_dim = int(len(b1[0]) / 2)
iou_mat = calculate_iou(b1, b2, dim=inferred_dim)
# feature similarity
if feature_similarity_beta is not None:
# get features
f1 = [t.feature for t in trackers]
f2 = [d.feature for d in detections]
if _sequence_has_none(f1) or _sequence_has_none(f2):
# fallback to pure IOU due to missing features
apt_mat = iou_mat
else:
sim_mat = feature_similarity_fn(f1, f2)
sim_mat = feature_similarity_beta + (
1 - feature_similarity_beta) * sim_mat
# combined aptitude
apt_mat = np.multiply(iou_mat, sim_mat)
else:
apt_mat = iou_mat
cost_mat = -1.0 * apt_mat
return cost_mat, iou_mat
def test_ious_large():
# benchmarking test
t = current_milli_time()
for _ in range(100):
b1 = np.random.randn(30, 4)
b2 = np.random.randn(20, 4)
iou = calculate_iou(b1, b2, dim=2)
assert iou.shape == (30, 20)
tdiff = current_milli_time() - t
assert tdiff < 100