2martens/masterthesis
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Fixed application of entropy threshold

Browse Source 
Signed-off-by: Jim Martens <github@2martens.de>
This commit is contained in:
Jim Martens
2019-08-29 13:31:13 +02:00
parent c403560b99
commit cf132a89f0
1 changed files with 16 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

28
src/twomartens/masterthesis/ssd.py
View File

@ -411,11 +411,11 @@ def _apply_entropy_threshold(observations: Sequence[np.ndarray], entropy_thresho
final_observations.append(observations[i]) final_observations.append(observations[i])
continue continue
filtered_image_observations = observations[observations[i][-1] < entropy_threshold] filtered_image_observations = observations[observations[i][:, -1] < entropy_threshold]
final_image_observations = np.copy(filtered_image_observations[:, :, -8:-1]) final_image_observations = np.copy(filtered_image_observations[:, -8:-1])
final_image_observations[:, :, 0] = np.argmax(filtered_image_observations[:, :, :-5], axis=-1) final_image_observations[:, 0] = np.argmax(filtered_image_observations[:, :-5], axis=-1)
final_image_observations[:, :, 1] = np.amax(filtered_image_observations[:, :, :-5], axis=-1) final_image_observations[:, 1] = np.amax(filtered_image_observations[:, :-5], axis=-1)
final_image_observations[:, :, 2] = filtered_image_observations[:, :, -1] final_image_observations[:, 2] = filtered_image_observations[:, -1]
final_observations.append(final_image_observations) final_observations.append(final_image_observations)
return final_observations return final_observations
@ -479,7 +479,7 @@ def _get_observations(detections: Sequence[np.ndarray]) -> List[np.ndarray]:
if not len(overlap_detections[0]): if not len(overlap_detections[0]):
continue continue
observation_set = np.unique(overlap_detections, axis=0) observation_set = np.unique(overlap_detections, axis=0)
for k in overlap_detections: for k in overlap_detections[0]:
# check if box was already removed from observation, then skip # check if box was already removed from observation, then skip
if k not in observation_set: if k not in observation_set:
continue continue
@ -488,8 +488,8 @@ def _get_observations(detections: Sequence[np.ndarray]) -> List[np.ndarray]:
# detection # detection
second_overlaps = overlaps[k] second_overlaps = overlaps[k]
second_detections = np.unique(np.nonzero(second_overlaps >= 0.95), axis=0) second_detections = np.unique(np.nonzero(second_overlaps >= 0.95), axis=0)
difference = _set_difference(observation_set, second_detections) difference = np.setdiff1d(observation_set, second_detections, assume_unique=True)
observation_set = _set_difference(observation_set, difference) observation_set = np.setdiff1d(observation_set, difference)
if used_boxes is None: if used_boxes is None:
used_boxes = observation_set used_boxes = observation_set
@ -522,7 +522,11 @@ def _set_difference(first_array: np.ndarray, second_array: np.ndarray) -> np.nda
Returns: Returns:
set difference between first_array and second_array set difference between first_array and second_array
""" """
dims = np.maximum(second_array.max(axis=0), max2 = second_array.max(axis=0)
first_array.max(axis=0)) + 1 max1 = first_array.max(axis=0)
return second_array[~np.in1d(np.ravel_multi_index(second_array.T, dims), dims = np.maximum(max2,
np.ravel_multi_index(first_array.T, dims))] max1) + 1
rmi2 = np.ravel_multi_index(second_array.T, dims)
rmi1 = np.ravel_multi_index(first_array.T, dims)
in1d = np.in1d(rmi2, rmi1)
return second_array[~in1d]