def pts2bbox(pts, debug=True, vis=False):
'''
convert a set of 2d points to a bounding box
parameter:
pts: 2 x N numpy array, N should >= 2
return:
bbox: 1 x 4 numpy array, TLBR format
'''
if debug:
assert is2dptsarray(pts) or is2dptsarray_occlusion(
pts
), 'the input points should have shape: 2 or 3 x num_pts vs %d x %s' % (
pts.shape[0], pts.shape[1])
assert pts.shape[
1] >= 2, 'number of points should be larger or equal than 2'
bbox = np.zeros((1, 4), dtype='float32')
bbox[0, 0] = np.min(pts[0, :]) # x coordinate of left top point
bbox[0, 1] = np.min(pts[1, :]) # y coordinate of left top point
bbox[0, 2] = np.max(pts[0, :]) # x coordinate of bottom right point
bbox[0, 3] = np.max(pts[1, :]) # y coordinate of bottom right point
if vis:
fig = plt.figure()
pts = imagecoor2cartesian(pts)
plt.scatter(pts[0, :], pts[1, :], color='r')
plt.scatter(
bbox[0, 0], -bbox[0, 1], color='b'
) # -1 is to convert the coordinate from image to cartesian
plt.scatter(bbox[0, 2], -bbox[0, 3], color='b')
plt.show()
plt.close(fig)
return bbox
def test_is2dptsarray():
pts = np.array([[1, 3], [2, 4]])
assert is2dptsarray(pts)
pts = np.array([[1, 3, 5], [2, 4, 3]])
assert is2dptsarray(pts)
pts = np.random.rand(2, 0)
assert is2dptsarray(pts)
pts = np.random.rand(3, 0)
assert not is2dptsarray(pts)
pts = [[1, 3], [2, 4]]
assert not is2dptsarray(pts)
pts = np.random.rand(1, 0)
assert not is2dptsarray(pts)
def pts_conversion_back_bbox(pts_array, bbox, debug=True):
'''
convert pts in the cropped image to the pts in the original image
parameters:
bbox: 1 X 4 numpy array, TLBR or TLWH format
pts_array: 2(3) x N numpy array, N should >= 1
'''
if debug:
assert is2dptsarray(pts_array) or is2dptsarray_occlusion(
pts_array
), 'the input points should have shape: 2 or 3 x num_pts vs %d x %s' % (
pts_array.shape[0], pts_array.shape[1])
assert bboxcheck(bbox), 'the input bounding box is not correct'
pts_array[0, :] = pts_array[0, :] + bbox[0, 0]
pts_array[1, :] = pts_array[1, :] + bbox[0, 1]
return pts_array
def facial_landmark_evaluation(pred_dict_all,
anno_dict,
num_pts,
error_threshold,
normalization_ced=True,
normalization_vec=False,
covariance=True,
display_list=None,
debug=True,
vis=False,
save=True,
save_path=None):
'''
evaluate the performance of facial landmark detection
parameter:
pred_dict_all: a dictionary for all basline methods. Each key is the method name and the value is corresponding prediction dictionary,
which keys are the image path and values are 2 x N prediction results
anno_dict: a dictionary which keys are the image path and values are 2 x N annotation results
num_pts: number of points
vis: determine if visualizing the pck curve
save: determine if saving the visualization results
save_path: a directory to save all the results
visualization:
1. 2d pck curve (total and point specific) for all points for all methods
2. point error vector (total and point specific) for all points and for all methods
3. mean square error
return:
metrics_all: a list of list to have detailed metrics over all methods
ptswise_mse: a list of list to have average MSE over all key-points for all methods
'''
num_methods = len(pred_dict_all)
if debug:
assert isdict(pred_dict_all) and num_methods > 0 and all(
isdict(pred_dict) for pred_dict in
pred_dict_all.values()), 'predictions result format is not correct'
assert isdict(anno_dict), 'annotation result format is not correct'
assert ispositiveinteger(num_pts), 'number of points is not correct'
assert isscalar(error_threshold), 'error threshold is not correct'
assert islogical(normalization_ced) and islogical(
normalization_vec), 'normalization flag is not correct'
if display_list is not None:
assert len(
display_list
) == num_methods, 'display list is not correct %d vs %d' % (
len(display_list), num_methods)
num_images = len(pred_dict_all.values()[0])
if debug:
assert num_images > 0, 'the predictions are empty'
assert num_images == len(
anno_dict
), 'number of images is not equal to number of annotations: %d vs %d' % (
num_images, len(anno_dict))
assert all(
num_images == len(pred_dict)
for pred_dict in pred_dict_all.values()
), 'number of images in results from different methods are not equal'
# calculate normalized mean error for each single image based on point-to-point Euclidean distance normalized by the bounding box size
# calculate point error vector for each single image based on error vector normalized by the bounding box size
normed_mean_error_dict = dict()
normed_mean_error_pts_specific_dict = dict()
normed_mean_error_pts_specific_valid_dict = dict()
pts_error_vec_dict = dict()
pts_error_vec_pts_specific_dict = dict()
for method_name, pred_dict in pred_dict_all.items():
normed_mean_error_total = np.zeros((num_images, ), dtype='float32')
normed_mean_error_pts_specific = np.zeros((num_images, num_pts),
dtype='float32')
normed_mean_error_pts_specific_valid = np.zeros((num_images, num_pts),
dtype='bool')
pts_error_vec = np.zeros((num_images, 2), dtype='float32')
pts_error_vec_pts_specific = np.zeros((num_images, 2, num_pts),
dtype='float32')
count = 0
count_skip_num_images = 0 # it's possible that no annotation exists on some images, than no error should be counted for those images, we count the number of those images
for image_path, pts_prediction in pred_dict.items():
_, filename, _ = fileparts(image_path)
pts_anno = anno_dict[filename] # 2 x N annotation
pts_keep_index = range(num_pts)
# to avoid list object type, do conversion here
if islist(pts_anno):
pts_anno = np.asarray(pts_anno)
if islist(pts_prediction):
pts_prediction = np.asarray(pts_prediction)
if debug:
assert (
is2dptsarray(pts_anno) or is2dptsarray_occlusion(pts_anno)
) and pts_anno.shape[
1] == num_pts, 'shape of annotations is not correct (%d x %d) vs (%d x %d)' % (
2, num_pts, pts_anno.shape[0], pts_anno.shape[1])
# if the annotation has 3 channels (include extra occlusion channel, we keep only the points with annotations)
# occlusion: -1 -> not annotated, 0 -> invisible, 1 -> visible, we keep both visible and invisible points
if pts_anno.shape[0] == 3:
pts_keep_index = np.where(
np.logical_or(pts_anno[2, :] == 1,
pts_anno[2, :] == 0))[0].tolist()
if len(pts_keep_index
) <= 0: # if no point is annotated in current image
count_skip_num_images += 1
continue
pts_anno = pts_anno[0:2, pts_keep_index]
pts_prediction = pts_prediction[:, pts_keep_index]
# to avoid the point location includes the score or occlusion channel, only take the first two channels here
if pts_prediction.shape[0] == 3 or pts_prediction.shape[0] == 4:
pts_prediction = pts_prediction[0:2, :]
num_pts_tmp = len(pts_keep_index)
if debug:
assert pts_anno.shape[
1] <= num_pts, 'number of points is not correct: %d vs %d' % (
pts_anno.shape[1], num_pts)
assert pts_anno.shape == pts_prediction.shape, 'shape of annotations and predictions are not the same {} vs {}'.format(
print_np_shape(pts_anno, debug=debug),
print_np_shape(pts_prediction, debug=debug))
# print 'number of points to keep is %d' % num_pts_tmp
# calculate bbox for normalization
if normalization_ced or normalization_vec:
assert len(
pts_keep_index
) == num_pts, 'some points are not annotated. Normalization on PCK curve is not allowed.'
bbox_anno = pts2bbox(pts_anno, debug=debug) # 1 x 4
bbox_TLWH = bbox_TLBR2TLWH(bbox_anno, debug=debug) # 1 x 4
bbox_size = math.sqrt(bbox_TLWH[0, 2] *
bbox_TLWH[0, 3]) # scalar
# calculate normalized error for all points
normed_mean_error, _ = pts_euclidean(pts_prediction,
pts_anno,
debug=debug) # scalar
if normalization_ced:
normed_mean_error /= bbox_size
normed_mean_error_total[count] = normed_mean_error
if normed_mean_error == 0:
print pts_prediction
print pts_anno
# calculate normalized error point specifically
for pts_index in xrange(num_pts):
if pts_index in pts_keep_index: # if current point not annotated in current image, just keep 0
normed_mean_error_pts_specific_valid[count,
pts_index] = True
else:
continue
pts_index_from_keep_list = pts_keep_index.index(pts_index)
pts_prediction_tmp = np.reshape(
pts_prediction[:, pts_index_from_keep_list], (2, 1))
pts_anno_tmp = np.reshape(
pts_anno[:, pts_index_from_keep_list], (2, 1))
normed_mean_error_pts_specifc_tmp, _ = pts_euclidean(
pts_prediction_tmp, pts_anno_tmp, debug=debug)
if normalization_ced:
normed_mean_error_pts_specifc_tmp /= bbox_size
normed_mean_error_pts_specific[
count, pts_index] = normed_mean_error_pts_specifc_tmp
# calculate the point error vector
error_vector = pts_prediction - pts_anno # 2 x num_pts_tmp
if normalization_vec:
error_vector /= bbox_size
pts_error_vec_pts_specific[
count, :, pts_keep_index] = np.transpose(error_vector)
pts_error_vec[count, :] = np.sum(error_vector,
axis=1) / num_pts_tmp
count += 1
assert count + count_skip_num_images == num_images, 'all cells in the array must be filled %d vs %d' % (
count + count_skip_num_images, num_images)
# print normed_mean_error_total
# time.sleep(1000)
# save results to dictionary
normed_mean_error_dict[method_name] = normed_mean_error_total[:count]
normed_mean_error_pts_specific_dict[
method_name] = normed_mean_error_pts_specific[:count, :]
normed_mean_error_pts_specific_valid_dict[
method_name] = normed_mean_error_pts_specific_valid[:count, :]
pts_error_vec_dict[method_name] = np.transpose(
pts_error_vec[:count, :]) # 2 x num_images
pts_error_vec_pts_specific_dict[
method_name] = pts_error_vec_pts_specific[:count, :, :]
# calculate mean value
if mse:
mse_value = dict(
) # dictionary to record all average MSE for different methods
mse_dict = dict(
) # dictionary to record all point-wise MSE for different keypoints
for method_name, error_array in normed_mean_error_dict.items():
mse_value[method_name] = np.mean(error_array)
else:
mse_value = None
# visualize the ced (cumulative error distribution curve)
print('visualizing pck curve....\n')
pck_savedir = os.path.join(save_path, 'pck')
mkdir_if_missing(pck_savedir)
pck_savepath = os.path.join(pck_savedir, 'pck_curve_overall.png')
table_savedir = os.path.join(save_path, 'metrics')
mkdir_if_missing(table_savedir)
table_savepath = os.path.join(table_savedir,
'detailed_metrics_overall.txt')
_, metrics_all = visualize_ced(normed_mean_error_dict,
error_threshold=error_threshold,
normalized=normalization_ced,
truncated_list=truncated_list,
title='2D PCK curve (all %d points)' %
num_pts,
display_list=display_list,
debug=debug,
vis=vis,
save=save,
pck_savepath=pck_savepath,
table_savepath=table_savepath)
metrics_title = ['Method Name / Point Index']
ptswise_mse_table = [[
normed_mean_error_pts_specific_dict.keys()[index_tmp]
] for index_tmp in xrange(num_methods)]
for pts_index in xrange(num_pts):
metrics_title.append(str(pts_index + 1))
normed_mean_error_dict_tmp = dict()
for method_name, error_array in normed_mean_error_pts_specific_dict.items(
):
normed_mean_error_pts_specific_valid_temp = normed_mean_error_pts_specific_valid_dict[
method_name]
# Some points at certain images might not be annotated. When calculating MSE for these specific point, we remove those images to avoid "false" mean average error
valid_array_per_pts_per_method = np.where(
normed_mean_error_pts_specific_valid_temp[:, pts_index] ==
True)[0].tolist()
error_array_per_pts = error_array[:, pts_index]
error_array_per_pts = error_array_per_pts[
valid_array_per_pts_per_method]
num_image_tmp = len(valid_array_per_pts_per_method)
normed_mean_error_dict_tmp[method_name] = np.reshape(
error_array_per_pts, (num_image_tmp, ))
pck_savepath = os.path.join(pck_savedir,
'pck_curve_pts_%d.png' % (pts_index + 1))
table_savepath = os.path.join(
table_savedir, 'detailed_metrics_pts_%d.txt' % (pts_index + 1))
metrics_dict, _ = visualize_ced(normed_mean_error_dict_tmp,
error_threshold=error_threshold,
normalized=normalization_ced,
truncated_list=truncated_list,
display2terminal=False,
title='2D PCK curve for point %d' %
(pts_index + 1),
display_list=display_list,
debug=debug,
vis=vis,
save=save,
pck_savepath=pck_savepath,
table_savepath=table_savepath)
for method_index in range(num_methods):
method_name = normed_mean_error_pts_specific_dict.keys(
)[method_index]
ptswise_mse_table[method_index].append(
'%.1f' % metrics_dict[method_name]['MSE'])
# reorder the table
order_index_list = [
display_list.index(method_name_tmp)
for method_name_tmp in normed_mean_error_pts_specific_dict.keys()
]
order_index_list = [0] + [
order_index_tmp + 1 for order_index_tmp in order_index_list
]
# print table to terminal
ptswise_mse_table = list_reorder([metrics_title] + ptswise_mse_table,
order_index_list,
debug=debug)
table = AsciiTable(ptswise_mse_table)
print '\nprint point-wise average MSE'
print table.table
# save table to file
ptswise_savepath = os.path.join(table_savedir, 'pointwise_average_MSE.txt')
table_file = open(ptswise_savepath, 'w')
table_file.write(table.table)
table_file.close()
print '\nsave point-wise average MSE to %s' % ptswise_savepath
# visualize the error vector map
print('visualizing error vector distribution map....\n')
error_vec_save_dir = os.path.join(save_path, 'error_vec')
mkdir_if_missing(error_vec_save_dir)
savepath_tmp = os.path.join(error_vec_save_dir,
'error_vector_distribution_all.png')
visualize_pts(pts_error_vec_dict,
title='Point Error Vector Distribution (all %d points)' %
num_pts,
mse=mse,
mse_value=mse_value,
display_range=display_range,
display_list=display_list,
xlim=xlim,
ylim=ylim,
covariance=covariance,
debug=debug,
vis=vis,
save=save,
save_path=savepath_tmp)
for pts_index in xrange(num_pts):
pts_error_vec_pts_specific_dict_tmp = dict()
for method_name, error_vec_dict in pts_error_vec_pts_specific_dict.items(
):
pts_error_vec_pts_specific_valid = normed_mean_error_pts_specific_valid_dict[
method_name] # get valid flag
valid_image_index_per_pts = np.where(
pts_error_vec_pts_specific_valid[:, pts_index] == True
)[0].tolist(
) # get images where the points with current index are annotated
pts_error_vec_pts_specific_dict_tmp[method_name] = np.transpose(
error_vec_dict[valid_image_index_per_pts, :,
pts_index]) # 2 x num_images
savepath_tmp = os.path.join(
error_vec_save_dir,
'error_vector_distribution_pts_%d.png' % (pts_index + 1))
if mse:
mse_dict_tmp = visualize_pts(
pts_error_vec_pts_specific_dict_tmp,
title='Point Error Vector Distribution for Point %d' %
(pts_index + 1),
mse=mse,
display_range=display_range,
display_list=display_list,
xlim=xlim,
ylim=ylim,
covariance=covariance,
debug=debug,
vis=vis,
save=save,
save_path=savepath_tmp)
mse_best = min(mse_dict_tmp.values())
mse_single = dict()
mse_single['mse'] = mse_best
mse_single['num_images'] = len(
valid_image_index_per_pts
) # assume number of valid images is equal for all methods
mse_dict[pts_index] = mse_single
else:
visualize_pts(
pts_error_vec_pts_specific_dict_tmp,
title='Point Error Vector Distribution for Point %d' %
(pts_index + 1),
mse=mse,
display_range=display_range,
display_list=display_list,
xlim=xlim,
ylim=ylim,
covariance=covariance,
debug=debug,
vis=vis,
save=save,
save_path=savepath_tmp)
# save mse to json file for further use
if mse:
json_path = os.path.join(save_path, 'mse_pts.json')
with open(json_path, 'w') as file:
print('save mse for all keypoings to {}'.format(json_path))
json.dump(mse_dict, file)
file.close()
print('\ndone!!!!!\n')
return metrics_all, ptswise_mse_table