def test_calc_heatmap_mixed_args(self):
det = data_holders.BBoxDetInst([0.1], [15, 20.8, 45, 60.6], pos_prob=1)
expected_map = np.zeros((100, 100))
expected_map[21:61, 15:46] = 1
expected_map[20, 15:46] = 0.2
expected_map[61, 15:46] = 0.6
heatmap = det.calc_heatmap((100, 100))
self.assertNPClose(expected_map, heatmap)
def test_calc_heatmap_outsize_bounds(self):
det = data_holders.BBoxDetInst([0.1], [-15.5, 80.8, 45.3, 160.6], pos_prob=1)
expected_map = np.zeros((100, 100))
expected_map[81:100, 0:46] = 1
expected_map[81:100, 46] = 0.3
expected_map[80, 0:46] = 0.2
expected_map[80, 46] = 0.3 * 0.2
heatmap = det.calc_heatmap((100, 100))
self.assertNPClose(expected_map, heatmap)
def test_calc_heatmap_float_args(self):
det = data_holders.BBoxDetInst([0.1], [15.5, 20.8, 45.3, 60.6], pos_prob=1)
expected_map = np.zeros((100, 100))
expected_map[21:61, 16:46] = 1
expected_map[21:61, 15] = 0.5
expected_map[21:61, 46] = 0.3
expected_map[20, 16:46] = 0.2
expected_map[61, 16:46] = 0.6
expected_map[20, 15] = 0.5 * 0.2
expected_map[61, 15] = 0.5 * 0.6
expected_map[20, 46] = 0.3 * 0.2
expected_map[61, 46] = 0.3 * 0.6
heatmap = det.calc_heatmap((100, 100))
self.assertNPClose(expected_map, heatmap)
def __iter__(self):
"""
Generate DetectionInstances for a given image.
DetectionInstances will be of sub-class BBoxDetInst or PBoxDetInst.
BBoxDetInst is DetectionInstance for standard bounding box detection.
PBoxDetInst is DetectionInstance for probabilistic bounding box detection.
:param img_dets: list of detections given as dictionaries.
Individual detection dictionaries have the keys:
'img_size': (height x width)
'img_num': int identifying which image the detection is a part of
'label_probs': full list of probabilities that the detection is describing each of the classes
'bbox': coordinates of the bounding box corners [left, top, right, bottom]
'covars': covariances for the top-left and bottom-right corners respectively.
Each with format [[xx, xy], [yx, yy]]. Covariances must be positive semi-definite
or all zeros (regular BBox).
:param class_mapping: A pair of lists of indexes, the first to our class list, and the second to theirs
:param num_classes: The number of classes to expect
:param img_idx: The current image index, for error reporting
:param sequence_name: The current image name, for error reporting
:return: generator of DetectionInstances
"""
# Handle no detections for the image
for det_idx, det in enumerate(self._img_dets):
if 'label_probs' not in det:
raise KeyError(make_error_msg("missing key \'label_probs\'", self._sequence_name, self._img_idx, det_idx))
if 'bbox' not in det:
raise KeyError(make_error_msg("missing key \'bbox\'", self._sequence_name, self._img_idx, det_idx))
if len(det['label_probs']) != self._num_classes:
raise KeyError(make_error_msg("The number of class probabilities doesn't match the number of classes",
self._sequence_name, self._img_idx, det_idx))
if len(det['bbox']) != 4:
raise ValueError(make_error_msg("The bounding box must contain exactly 4 entries",
self._sequence_name, self._img_idx, det_idx))
if det['bbox'][2] < det['bbox'][0]:
raise ValueError(make_error_msg("The x1 coordinate must be less than the x2 coordinate",
self._sequence_name, self._img_idx, det_idx))
if det['bbox'][3] < det['bbox'][1]:
raise ValueError(make_error_msg("The y1 coordinate must be less than the y2 coordinate",
self._sequence_name, self._img_idx, det_idx))
# Use numpy list indexing to move specific indexes from the submission
label_probs = np.zeros(len(rvc1_class_list.CLASSES), dtype=np.float32)
label_probs[self._class_mapping[0]] = np.array(det['label_probs'])[self._class_mapping[1]]
total_prob = np.sum(label_probs)
if total_prob > 0.5: # Arbitrary theshold for classes we care about.
# Normalize the label probability
if total_prob > 1:
label_probs /= total_prob
if 'covars' not in det or det['covars'] == [[[0, 0], [0, 0]], [[0, 0], [0, 0]]] or self._override_cov == 0:
yield data_holders.BBoxDetInst(
class_list=label_probs,
box=det['bbox']
)
else:
if self._override_cov is not None:
covars = np.array([[[self._override_cov, 0], [0, self._override_cov]],
[[self._override_cov, 0], [0, self._override_cov]]])
else:
covars = np.array(det['covars'])
if covars.shape != (2, 2, 2):
raise ValueError(make_error_msg("Key 'covars' must contain 2 2x2 matrices",
self._sequence_name, self._img_idx, det_idx))
if not np.allclose(covars.transpose((0, 2, 1)), covars):
raise ValueError(make_error_msg("Given covariances are not symmetric",
self._sequence_name, self._img_idx, det_idx))
if not is_positive_semi_definite(covars[0]):
raise ValueError(make_error_msg("The upper-left covariance is not positive semi-definite",
self._sequence_name, self._img_idx, det_idx))
if not is_positive_semi_definite(covars[1]):
raise ValueError(make_error_msg("The lower-right covariance is not positive semi-definite",
self._sequence_name, self._img_idx, det_idx))
yield data_holders.PBoxDetInst(
class_list=label_probs,
box=det['bbox'],
covs=det['covars']
)
def test_calc_heatmap_int_args(self):
det = data_holders.BBoxDetInst([0.1], [15, 20, 45, 60], pos_prob=1)
expected_map = np.zeros((100, 100))
expected_map[20:61, 15:46] = 1
heatmap = det.calc_heatmap((100, 100))
self.assertNPEqual(expected_map, heatmap)