def call(self, x, mask=None):
"""
Return an anchor box tensor based on the shape of the input tensor.
The logic implemented here is identical to the logic in the module `ssd_box_encode_decode_utils.py`.
Note that this tensor does not participate in any graph computations at runtime. It is being created
as a constant once during graph creation and is just being output along with the rest of the model output
during runtime. Because of this, all logic is implemented as Numpy array operations and it is sufficient
to convert the resulting Numpy array into a Keras tensor at the very end before outputting it.
Arguments:
x (tensor): 4D tensor of shape `(batch, channels, height, width)` if `dim_ordering = 'th'`
or `(batch, height, width, channels)` if `dim_ordering = 'tf'`. The input for this
layer must be the output of the localization predictor layer.
"""
# Compute box width and height for each aspect ratio
# The shorter side of the image will be used to compute `w` and `h` using `scale` and `aspect_ratios`.
size = min(self.img_height, self.img_width)
# Compute the box widths and and heights for all aspect ratios
wh_list = []
for ar in self.aspect_ratios:
if (ar == 1):
# Compute the regular anchor box for aspect ratio 1.
box_height = box_width = self.this_scale * size
wh_list.append((box_width, box_height))
if self.two_boxes_for_ar1:
# Compute one slightly larger version using the geometric mean of this scale value and the next.
box_height = box_width = np.sqrt(self.this_scale * self.next_scale) * size
wh_list.append((box_width, box_height))
else:
box_height = self.this_scale * size / np.sqrt(ar)
box_width = self.this_scale * size * np.sqrt(ar)
wh_list.append((box_width, box_height))
wh_list = np.array(wh_list)
# We need the shape of the input tensor
if K.common.image_dim_ordering() == 'tf':
batch_size, feature_map_height, feature_map_width, feature_map_channels = x._keras_shape
# Not yet relevant since TensorFlow is the only supported backend right now,
# but it can't harm to have this in here for the future
else:
batch_size, feature_map_channels, feature_map_height, feature_map_width = x._keras_shape
# Compute the grid of box center points. They are identical for all aspect ratios.
# Compute the step sizes, i.e. how far apart the anchor box center points will be vertically and horizontally.
if (self.this_steps is None):
step_height = self.img_height / feature_map_height
step_width = self.img_width / feature_map_width
else:
if isinstance(self.this_steps, (list, tuple)) and (len(self.this_steps) == 2):
step_height = self.this_steps[0]
step_width = self.this_steps[1]
elif isinstance(self.this_steps, (int, float)):
step_height = self.this_steps
step_width = self.this_steps
# Compute the offsets, i.e. at what pixel values the first anchor box center point will be
# from the top and from the left of the image.
if self.this_offsets is None:
offset_height = 0.5
offset_width = 0.5
else:
if isinstance(self.this_offsets, (list, tuple)) and (len(self.this_offsets) == 2):
offset_height = self.this_offsets[0]
offset_width = self.this_offsets[1]
elif isinstance(self.this_offsets, (int, float)):
offset_height = self.this_offsets
offset_width = self.this_offsets
# Now that we have the offsets and step sizes, compute the grid of anchor box center points.
cy = np.linspace(offset_height * step_height,
(offset_height + feature_map_height - 1) * step_height, feature_map_height)
cx = np.linspace(offset_width * step_width,
(offset_width + feature_map_width - 1) * step_width, feature_map_width)
cx_grid, cy_grid = np.meshgrid(cx, cy)
# This is necessary for np.tile() to do what we want further down
cx_grid = np.expand_dims(cx_grid, -1)
# This is necessary for np.tile() to do what we want further down
cy_grid = np.expand_dims(cy_grid, -1)
# Create a 4D tensor template of shape `(feature_map_height, feature_map_width, n_boxes, 4)`
# where the last dimension will contain `(cx, cy, w, h)`
boxes_tensor = np.zeros((feature_map_height, feature_map_width, self.n_boxes, 4))
boxes_tensor[:, :, :, 0] = np.tile(cx_grid, (1, 1, self.n_boxes)) # Set cx
boxes_tensor[:, :, :, 1] = np.tile(cy_grid, (1, 1, self.n_boxes)) # Set cy
boxes_tensor[:, :, :, 2] = wh_list[:, 0] # Set w
boxes_tensor[:, :, :, 3] = wh_list[:, 1] # Set h
# Convert `(cx, cy, w, h)` to `(xmin, xmax, ymin, ymax)`
boxes_tensor = convert_coordinates(boxes_tensor, start_index=0, conversion='centroids2corners')
# If `clip_boxes` is enabled, clip the coordinates to lie within the image boundaries
if self.clip_boxes:
x_coords = boxes_tensor[:, :, :, [0, 2]]
x_coords[x_coords >= self.img_width] = self.img_width - 1
x_coords[x_coords < 0] = 0
boxes_tensor[:, :, :, [0, 2]] = x_coords
y_coords = boxes_tensor[:, :, :, [1, 3]]
y_coords[y_coords >= self.img_height] = self.img_height - 1
y_coords[y_coords < 0] = 0
boxes_tensor[:, :, :, [1, 3]] = y_coords
# If `normalize_coords` is enabled, normalize the coordinates to be within [0,1]
if self.normalize_coords:
boxes_tensor[:, :, :, [0, 2]] /= self.img_width
boxes_tensor[:, :, :, [1, 3]] /= self.img_height
# TODO: Implement box limiting directly for `(cx, cy, w, h)` so that
# we don't have to unnecessarily convert back and forth.
if self.coords == 'centroids':
# Convert `(xmin, ymin, xmax, ymax)` back to `(cx, cy, w, h)`.
boxes_tensor = convert_coordinates(boxes_tensor, start_index=0, conversion='corners2centroids',
border_pixels='half')
elif self.coords == 'minmax':
# Convert `(xmin, ymin, xmax, ymax)` to `(xmin, xmax, ymin, ymax).
boxes_tensor = convert_coordinates(boxes_tensor, start_index=0, conversion='corners2minmax',
border_pixels='half')
# Create a tensor to contain the variances and append it to `boxes_tensor`. This tensor has the same shape
# as `boxes_tensor` and simply contains the same 4 variance values for every position in the last axis.
variances_tensor = np.zeros_like(
boxes_tensor) # Has shape `(feature_map_height, feature_map_width, n_boxes, 4)`
variances_tensor += self.variances # Long live broadcasting
# Now `boxes_tensor` becomes a tensor of shape `(feature_map_height, feature_map_width, n_boxes, 8)`
boxes_tensor = np.concatenate((boxes_tensor, variances_tensor), axis=-1)
# Now prepend one dimension to `boxes_tensor` to account for the batch size and tile it along
# The result will be a 5D tensor of shape `(batch_size, feature_map_height, feature_map_width, n_boxes, 8)`
boxes_tensor = np.expand_dims(boxes_tensor, axis=0)
boxes_tensor = K.tile(K.constant(boxes_tensor, dtype='float32'), (K.shape(x)[0], 1, 1, 1, 1))
return boxes_tensor
def generate_anchor_boxes_for_layer(self,
feature_map_size,
aspect_ratios,
this_scale,
next_scale,
this_steps=None,
this_offsets=None,
diagnostics=False):
"""
Computes an array of the spatial positions and sizes of the anchor boxes for one predictor layer
of size `feature_map_size == [feature_map_height, feature_map_width]`.
Arguments:
feature_map_size (tuple): A list or tuple `[feature_map_height, feature_map_width]` with the spatial
dimensions of the feature map for which to generate the anchor boxes.
aspect_ratios (list): A list of floats, the aspect ratios for which anchor boxes are to be generated.
All list elements must be unique.
this_scale (float): A float in [0, 1], the scaling factor for the size of the generate anchor boxes
as a fraction of the shorter side of the input image.
next_scale (float): A float in [0, 1], the next larger scaling factor. Only relevant if
`self.two_boxes_for_ar1 == True`.
diagnostics (bool, optional): If true, the following additional outputs will be returned:
1) A list of the center point `x` and `y` coordinates for each spatial location.
2) A list containing `(width, height)` for each box aspect ratio.
3) A tuple containing `(step_height, step_width)`
4) A tuple containing `(offset_height, offset_width)`
This information can be useful to understand in just a few numbers what the generated grid of
anchor boxes actually looks like, i.e. how large the different boxes are and how dense
their spatial distribution is, in order to determine whether the box grid covers the input images
appropriately and whether the box sizes are appropriate to fit the sizes of the objects
to be detected.
Returns:
A 4D Numpy tensor of shape `(feature_map_height, feature_map_width, n_boxes_per_cell, 4)` where the
last dimension contains `(xmin, xmax, ymin, ymax)` for each anchor box in each cell of the feature map.
"""
# Compute box width and height for each aspect ratio.
# The shorter side of the image will be used to compute `w` and `h` using `scale` and `aspect_ratios`.
size = min(self.img_height, self.img_width)
# Compute the box widths and and heights for all aspect ratios
wh_list = []
for ar in aspect_ratios:
if (ar == 1):
# Compute the regular anchor box for aspect ratio 1.
box_height = box_width = this_scale * size
wh_list.append((box_width, box_height))
if self.two_boxes_for_ar1:
# Compute one slightly larger version using the geometric mean of this scale value and the next.
box_height = box_width = np.sqrt(this_scale * next_scale) * size
wh_list.append((box_width, box_height))
else:
box_width = this_scale * size * np.sqrt(ar)
box_height = this_scale * size / np.sqrt(ar)
wh_list.append((box_width, box_height))
wh_list = np.array(wh_list)
n_boxes = len(wh_list)
# Compute the grid of box center points. They are identical for all aspect ratios.
# Compute the step sizes, i.e. how far apart the anchor box center points will be vertically and horizontally.
if (this_steps is None):
step_height = self.img_height / feature_map_size[0]
step_width = self.img_width / feature_map_size[1]
else:
if isinstance(this_steps, (list, tuple)) and (len(this_steps) == 2):
step_height = this_steps[0]
step_width = this_steps[1]
elif isinstance(this_steps, (int, float)):
step_height = this_steps
step_width = this_steps
# Compute the offsets, i.e. at what pixel values the first anchor box center point
# will be from the top and from the left of the image.
if (this_offsets is None):
offset_height = 0.5
offset_width = 0.5
else:
if isinstance(this_offsets, (list, tuple)) and (len(this_offsets) == 2):
offset_height = this_offsets[0]
offset_width = this_offsets[1]
elif isinstance(this_offsets, (int, float)):
offset_height = this_offsets
offset_width = this_offsets
# Now that we have the offsets and step sizes, compute the grid of anchor box center points.
cy = np.linspace(offset_height * step_height, (offset_height + feature_map_size[0] - 1) * step_height,
feature_map_size[0])
cx = np.linspace(offset_width * step_width, (offset_width + feature_map_size[1] - 1) * step_width,
feature_map_size[1])
cx_grid, cy_grid = np.meshgrid(cx, cy)
cx_grid = np.expand_dims(cx_grid, -1) # This is necessary for np.tile() to do what we want further down
cy_grid = np.expand_dims(cy_grid, -1) # This is necessary for np.tile() to do what we want further down
# Create a 4D tensor template of shape `(feature_map_height, feature_map_width, n_boxes, 4)`
# where the last dimension will contain `(cx, cy, w, h)`
boxes_tensor = np.zeros((feature_map_size[0], feature_map_size[1], n_boxes, 4))
boxes_tensor[:, :, :, 0] = np.tile(cx_grid, (1, 1, n_boxes)) # Set cx
boxes_tensor[:, :, :, 1] = np.tile(cy_grid, (1, 1, n_boxes)) # Set cy
boxes_tensor[:, :, :, 2] = wh_list[:, 0] # Set w
boxes_tensor[:, :, :, 3] = wh_list[:, 1] # Set h
# Convert `(cx, cy, w, h)` to `(xmin, ymin, xmax, ymax)`
boxes_tensor = convert_coordinates(boxes_tensor, start_index=0, conversion='centroids2corners')
# If `clip_boxes` is enabled, clip the coordinates to lie within the image boundaries
if self.clip_boxes:
x_coords = boxes_tensor[:, :, :, [0, 2]]
x_coords[x_coords >= self.img_width] = self.img_width - 1
x_coords[x_coords < 0] = 0
boxes_tensor[:, :, :, [0, 2]] = x_coords
y_coords = boxes_tensor[:, :, :, [1, 3]]
y_coords[y_coords >= self.img_height] = self.img_height - 1
y_coords[y_coords < 0] = 0
boxes_tensor[:, :, :, [1, 3]] = y_coords
# `normalize_coords` is enabled, normalize the coordinates to be within [0,1]
if self.normalize_coords:
boxes_tensor[:, :, :, [0, 2]] /= self.img_width
boxes_tensor[:, :, :, [1, 3]] /= self.img_height
# TODO: Implement box limiting directly for `(cx, cy, w, h)`
# so that we don't have to unnecessarily convert back and forth.
if self.coords == 'centroids':
# Convert `(xmin, ymin, xmax, ymax)` back to `(cx, cy, w, h)`.
boxes_tensor = convert_coordinates(boxes_tensor, start_index=0, conversion='corners2centroids',
border_pixels='half')
elif self.coords == 'minmax':
# Convert `(xmin, ymin, xmax, ymax)` to `(xmin, xmax, ymin, ymax).
boxes_tensor = convert_coordinates(boxes_tensor, start_index=0, conversion='corners2minmax',
border_pixels='half')
if diagnostics:
return boxes_tensor, (cy, cx), wh_list, (step_height, step_width), (offset_height, offset_width)
else:
return boxes_tensor
def decode_detections(y_pred,
confidence_thresh=0.01,
iou_threshold=0.45,
top_k=200,
input_coords='centroids',
normalize_coords=True,
img_height=None,
img_width=None,
border_pixels='half'):
"""
Convert model prediction output back to a format that contains only the positive
box predictions (i.e., the same format that `SSDInputEncoder` takes as input).
After the decoding, two steps of prediction filtering are performed for each loss
individually: First confidence thresholding, then greedy non-maximum suppression.
The filtering results for all classes are concatenated and the `top_k` overall
highest confidence results constitute the final predictions for a given batch item.
Arguments:
y_pred: The prediction output of the SSD model, expected to be a NumPY array of
shape `(batch_size, #boxes, #classes + 4 + 4 + 4)`, where `#boxes` is the
total number of boxes predicted by the model per image and the last axis
contains `[one-hot vector for the classes, 4 predicted coordinate offsets,
4 anchor box coordinate, 4 variances]`.
confidence_thresh: A float in [0, 1), the minimum classification confidence in
a specific positive class in order to be considered for the non-maximum
suppression stage for the respective class. A lower value will result in a
larger part of the selection process being done by the non-maximum
suppression stage, while a larger value will result in a larger part of
the selection process happening in the confidence thresholding stage.
iou_threshold: A float in [0, 1]. All boxes with a IoU similarity of greater
than `iou_threshold` with a locally maximal box will be removed from the
set of predictions for a given class, where `maximal` refers to the box
score.
top_k: The number of highest scoring predictions to be kept for each batch item
after the non-maximum suppression stage.
input_coords: The box coordinate format that the model outputs. Can be either
'centroids' for the format `(cx, cy, w, h)`, 'minmax' for the format
`(xmin, xmax, ymin, ymax)`, or 'corners' for the format
`(xmin, ymin, xmax, ymax)`.
normalize_coords: Set to `True` if the model outputs relative coordinates
(i.e., coordinates in [0,1]) and you wish to transform these relative
coordinates back to absolute coordinates. If the model outputs relative
coordinates, but you do not want to convert them back to absolute coordinates,
set this to `False`. Do not set this is `True` if the model already outputs
absolute coordinates, as that would result in incorrect coordinates. Requires
`img_height` and `img_width` if set to `True`.
img_height: The height of the input image. Only needed if `normalized_coords` is `True`.
img_width: The width of the input image. Only needed if `normalized_coords` is `True`.
border_pixels: How to treat the border pixels of the bounding boxes. Can be 'include',
'exclude' and 'half'. If 'include', the border pixels belong to the boxes.
If 'exclude', the border pixels do not belong to the boxes. If 'half', then one
of each of the two horizontal and vertical borders belong to the boxes, but not
the other.
Returns:
A python list of length `batch_size` where each list element represent the predicted
boxes for one image and contains NumPy array of shape `(boxes, 6)` where each row is
a box prediction for a non-background class for the respective image in the format
`[class_id, confidence, xmin, ymin, xmax, ymax]`.
"""
if normalize_coords and (img_height is None or img_width is None):
raise ValueError(
'If relative box coordinates are supposed to be converted to absolute'
'coordinates, the decoder needs the image size in order to decode the '
f'prediction, but `img_height` == {img_height} and `img_width` == '
f'{img_width}')
# Slice out the classes and the four offsets, throw away the anchor coordinates and variance,
# resulting in a tensor of shape `(batch, n_boxes, n_classes + 4 coordinates]`.
y_pred_decoded_raw = np.copy(y_pred[:, :, :-8])
if input_coords == 'centroids':
# exp(ln(w(pred)/w(anchor)) / w_variance * w_variance) == w(pred) / w(anchor)
# exp(ln(h(pred)/h(anchor)) / h_variance * h_variance) == h(pred) / h(anchor)
y_pred_decoded_raw[:, :, [-2, -1]] = np.exp(
y_pred_decoded_raw[:, :, [-2, -1]] * y_pred[:, :, [-2, -1]])
# (w(pred) / w(anchor)) * w(anchor) == w(pred)
# (h(pred) / h(anchor)) * h(anchor) == h(pred)
y_pred_decoded_raw[:, :, [-2, -1]] *= y_pred[:, :, [-6, -5]]
# (delta_cx(pred) / w(anchor) / cx_variance) * cx_variance * w(anchor) == delta_cx(pred)
# (delta_cy(pred) / h(anchor) / cy_variance) * cy_variance * h(anchor) == delta_cy(pred)
y_pred_decoded_raw[:, :,
[-4, -3]] *= y_pred[:, :,
[-4, -3]] * y_pred[:, :,
[-6, -5]]
# delta_cx(pred) + cx(anchor) == cx(pred)
# delta_cy(pred) + cy(anchor) == cy(pred)
y_pred_decoded_raw[:, :, [-4, -3]] += y_pred[:, :, [-8, -7]]
y_pred_decoded_raw = convert_coordinates(
y_pred_decoded_raw, start_index=-4, conversion='centroids2corners')
elif input_coords == 'minmax':
# delta(pred) / size(anchor) / variance * variance == delta(pred) / size(anchor)
# for all four coordinates, where 'size' refers to w or h, respectively
y_pred_decoded_raw[:, :, -4:] *= y_pred[:, :, -4:]
# delta_xmin(pred) / w(anchor) * w(anchor) == delta_xmin(pred)
# delta_xmax(pred) / w(anchor) * w(anchor) == delta_xmax(pred)
y_pred_decoded_raw[:, :, [-4, -3]] *= np.expand_dims(y_pred[:, :, -7] -
y_pred[:, :, -8],
axis=-1)
# delta_ymin(pred) / h(anchor) * h(anchor) == delta_ymin(pred)
# delta_ymax(pred) / h(anchor) * h(anchor) == delta_ymax(pred)
y_pred_decoded_raw[:, :, [-2, -1]] *= np.expand_dims(y_pred[:, :, -5] -
y_pred[:, :, -6],
axis=-1)
# delta(pred) + anchor == pred for all four coordinates
y_pred_decoded_raw[:, :, -4:] += y_pred[:, :, -8:-4]
y_pred_decoded_raw = convert_coordinates(y_pred_decoded_raw,
start_index=-4,
conversion='minmax2corners')
elif input_coords == 'corners':
# delta(pred) / size(anchor) / variance * variance == delta(pred) / size(anchor)
# for all four coordinates, where 'size' refers to w or h, respectively
y_pred_decoded_raw[:, :, -4:] *= y_pred[:, :, -4:]
# delta_xmin(pred) / w(anchor) * w(anchor) == delta_xmin(pred)
# delta_xmax(pred) / w(anchor) * w(anchor) == delta_xmax(pred)
y_pred_decoded_raw[:, :, [-4, -2]] *= np.expand_dims(y_pred[:, :, -6] -
y_pred[:, :, -8],
axis=-1)
# delta_ymin(pred) / h(anchor) * h(anchor) == delta_ymin(pred)
# delta_ymax(pred) / h(anchor) * h(anchor) == delta_ymax(pred)
y_pred_decoded_raw[:, :, [-3, -1]] *= np.expand_dims(y_pred[:, :, -5] -
y_pred[:, :, -7],
axis=-1)
# delta(pred) + anchor == pred for all four coordinates
y_pred_decoded_raw[:, :, -4:] += y_pred[:, :, -8:-4]
else:
raise ValueError(
"Unexpected value for `input_coords`. "
"Supported input coordinate formats are 'minmax', 'corners' and 'centroids'."
)
if normalize_coords:
y_pred_decoded_raw[:, :, [-4, -2]] *= img_width
y_pred_decoded_raw[:, :, [-3, -1]] *= img_height
# Apply confidence thresholding and non-maximum suppression per class.
n_classes = y_pred_decoded_raw.shape[-1] - 4
y_pred_decode = [] # Store the final predictions in this list.
for batch_item in y_pred_decoded_raw: # Batch item has the shape `[n_boxes, n_classes + 4 coords]`.
pred = [] # Store the final predictions for this batch item.
for class_id in range(
1, n_classes): # For each class except the background class.
# Keep only the confidences for that class,
# making this array of shape `[n_boxes, confidence + 4 coords]`
single_class = batch_item[:, [class_id, -4, -3, -2, -1]]
# Keep only those boxes with a confidence above the threshold.
threshold_met = single_class[single_class[:,
0] > confidence_thresh]
# If any boxes pass the threshold.
if threshold_met.shape[0] > 0:
# Perform non-maximum suppression.
maxima = _greedy_nms(threshold_met,
iou_threshold=iou_threshold,
coords='corners',
border_pixels=border_pixels)
# Expand the last dimension by one element to have room for the class ID.
# This is now an array of shape `[n_boxes, confidence + 4 coords + class_id]`.
maxima_output = np.zeros(
(maxima.shape[0], maxima.shape[1] + 1))
# Write the class ID to the first column
maxima_output[:, 0] = class_id
# Write the maxima to the other columns
maxima_output[:, 1:] = maxima
pred.append(maxima_output)
# If there are any predictions left.
if pred:
pred = np.concatenate(pred, axis=0)
# If we have more than `top_k` results left at this point,
# Otherwise there is nothing to filter.
if top_k != 'all' and pred.shape[0] > top_k:
# Get the indices of the `top_k` highest score maximum.
top_k_indices = np.argpartition(pred[:, 1],
kth=pred.shape[0] - top_k,
axis=0)[pred.shape[0] - top_k:]
pred = pred[top_k_indices]
else:
pred = np.array(pred)
y_pred_decode.append(pred)
return y_pred_decode
def __call__(self, ground_truth_labels, diagnostics=False):
"""
Converts ground truth bounding box data into a suitable format to train an SSD model.
Arguments:
ground_truth_labels (list): A python list of length `batch_size` that contains one 2D Numpy array
for each batch image. Each such array has `k` rows for the `k` ground truth bounding boxes belonging
to the respective image, and the data for each ground truth bounding box has the format
`(class_id, xmin, ymin, xmax, ymax)` (i.e. the 'corners' coordinate format), and `class_id` must be
an integer greater than 0 for all boxes as class ID 0 is reserved for the background class.
diagnostics (bool, optional): If `True`, not only the encoded ground truth tensor will be returned,
but also a copy of it with anchor box coordinates in place of the ground truth coordinates.
This can be very useful if you want to visualize which anchor boxes got matched to which ground truth
boxes.
Returns:
`y_encoded`, a 3D numpy array of shape `(batch_size, #boxes, #classes + 4 + 4 + 4)` that serves as the
ground truth label tensor for training, where `#boxes` is the total number of boxes predicted by the
model per image, and the classes are one-hot-encoded. The four elements after the class vecotrs in
the last axis are the box coordinates, the next four elements after that are just dummy elements, and
the last four elements are the variances.
"""
# Mapping to define which indices represent which coordinates in the ground truth.
class_id = 0
xmin = 1
ymin = 2
xmax = 3
ymax = 4
batch_size = len(ground_truth_labels)
##################################################################################
# Generate the template for y_encoded.
##################################################################################
y_encoded = self.generate_encoding_template(batch_size=batch_size, diagnostics=False)
##################################################################################
# Match ground truth boxes to anchor boxes.
##################################################################################
# Match the ground truth boxes to the anchor boxes. Every anchor box that does not have
# a ground truth match and for which the maximal IoU overlap with any ground truth box is less
# than or equal to `neg_iou_limit` will be a negative (background) box.
y_encoded[:, :, self.background_id] = 1 # All boxes are background boxes by default.
n_boxes = y_encoded.shape[1] # The total number of boxes that the model predicts per batch item
class_vectors = np.eye(self.n_classes) # An identity matrix that we'll use as one-hot class vectors
for i in range(batch_size): # For each batch item...
if ground_truth_labels[
i].size == 0: continue # If there is no ground truth for this batch item, there is nothing to match.
labels = ground_truth_labels[i].astype(np.float) # The labels for this batch item
# Check for degenerate ground truth bounding boxes before attempting any computations.
if np.any(labels[:, [xmax]] - labels[:, [xmin]] <= 0) or np.any(labels[:, [ymax]] - labels[:, [ymin]] <= 0):
raise DegenerateBoxError(
"SSDInputEncoder detected degenerate ground truth bounding boxes \
for batch item {} with bounding boxes {}, ".format(
i, labels) +
"i.e. bounding boxes where xmax <= xmin and/or ymax <= ymin. Degenerate ground truth " +
"bounding boxes will lead to NaN errors during the training.")
# Maybe normalize the box coordinates.
if self.normalize_coords:
labels[:, [ymin, ymax]] /= self.img_height # Normalize ymin and ymax relative to the image height
labels[:, [xmin, xmax]] /= self.img_width # Normalize xmin and xmax relative to the image width
# Maybe convert the box coordinate format.
if self.coords == 'centroids':
labels = convert_coordinates(labels, start_index=xmin, conversion='corners2centroids',
border_pixels=self.border_pixels)
elif self.coords == 'minmax':
labels = convert_coordinates(labels, start_index=xmin, conversion='corners2minmax')
classes_one_hot = class_vectors[labels[:, class_id].astype(
np.int)] # The one-hot class IDs for the ground truth boxes of this batch item
labels_one_hot = np.concatenate([classes_one_hot, labels[:, [xmin, ymin, xmax, ymax]]],
axis=-1) # The one-hot version of the labels for this batch item
# Compute the IoU similarities between all anchor boxes and all ground truth boxes for this batch item.
# This is a matrix of shape `(num_ground_truth_boxes, num_anchor_boxes)`.
similarities = iou(labels[:, [xmin, ymin, xmax, ymax]], y_encoded[i, :, -12:-8], coords=self.coords,
mode='outer_product', border_pixels=self.border_pixels)
# First: Do bipartite matching, i.e. match each ground truth box to the one anchor box with the highest IoU.
# This ensures that each ground truth box will have at least one good match.
# For each ground truth box, get the anchor box to match with it.
bipartite_matches = match_bipartite_greedy(weight_matrix=similarities)
# Write the ground truth data to the matched anchor boxes.
y_encoded[i, bipartite_matches, :-8] = labels_one_hot
# Set the columns of the matched anchor boxes to zero to indicate that they were matched.
similarities[:, bipartite_matches] = 0
# Second: Maybe do 'multi' matching, where each remaining anchor box will be matched to its most similar
# ground truth box with an IoU of at least `pos_iou_threshold`, or not matched if there is no
# such ground truth box.
if self.matching_type == 'multi':
# Get all matches that satisfy the IoU threshold.
matches = match_multi(weight_matrix=similarities, threshold=self.pos_iou_threshold)
# Write the ground truth data to the matched anchor boxes.
y_encoded[i, matches[1], :-8] = labels_one_hot[matches[0]]
# Set the columns of the matched anchor boxes to zero to indicate that they were matched.
similarities[:, matches[1]] = 0
# Third: Now after the matching is done, all negative (background) anchor boxes that have
# an IoU of `neg_iou_limit` or more with any ground truth box will be set to netral,
# i.e. they will no longer be background boxes. These anchors are "too close" to a
# ground truth box to be valid background boxes.
max_background_similarities = np.amax(similarities, axis=0)
neutral_boxes = np.nonzero(max_background_similarities >= self.neg_iou_limit)[0]
y_encoded[i, neutral_boxes, self.background_id] = 0
##################################################################################
# Convert box coordinates to anchor box offsets.
##################################################################################
if self.coords == 'centroids':
# cx(gt) - cx(anchor), cy(gt) - cy(anchor)
y_encoded[:, :, [-12, -11]] -= y_encoded[:, :, [-8, -7]]
# (cx(gt) - cx(anchor)) / w(anchor) / cx_variance, (cy(gt) - cy(anchor)) / h(anchor) / cy_variance
y_encoded[:, :, [-12, -11]] /= y_encoded[:, :, [-6, -5]] * y_encoded[:, :, [-4, -3]]
# w(gt) / w(anchor), h(gt) / h(anchor)
y_encoded[:, :, [-10, -9]] /= y_encoded[:, :, [-6, -5]]
# ln(w(gt) / w(anchor)) / w_variance, ln(h(gt) / h(anchor)) / h_variance (ln == natural logarithm)
y_encoded[:, :, [-10, -9]] = np.log(y_encoded[:, :, [-10, -9]]) / y_encoded[:, :, [-2, -1]]
elif self.coords == 'corners':
# (gt - anchor) for all four coordinates
y_encoded[:, :, -12:-8] -= y_encoded[:, :, -8:-4]
# (xmin(gt) - xmin(anchor)) / w(anchor), (xmax(gt) - xmax(anchor)) / w(anchor)
y_encoded[:, :, [-12, -10]] /= np.expand_dims(y_encoded[:, :, -6] - y_encoded[:, :, -8], axis=-1)
# (ymin(gt) - ymin(anchor)) / h(anchor), (ymax(gt) - ymax(anchor)) / h(anchor)
y_encoded[:, :, [-11, -9]] /= np.expand_dims(y_encoded[:, :, -5] - y_encoded[:, :, -7], axis=-1)
# (gt - anchor) / size(anchor) / variance for all four coordinates,
# where 'size' refers to w and h respectively
y_encoded[:, :, -12:-8] /= y_encoded[:, :, -4:]
elif self.coords == 'minmax':
# (gt - anchor) for all four coordinates
y_encoded[:, :, -12:-8] -= y_encoded[:, :, -8:-4]
# (xmin(gt) - xmin(anchor)) / w(anchor), (xmax(gt) - xmax(anchor)) / w(anchor)
y_encoded[:, :, [-12, -11]] /= np.expand_dims(y_encoded[:, :, -7] - y_encoded[:, :, -8], axis=-1)
# (ymin(gt) - ymin(anchor)) / h(anchor), (ymax(gt) - ymax(anchor)) / h(anchor)
y_encoded[:, :, [-10, -9]] /= np.expand_dims(y_encoded[:, :, -5] - y_encoded[:, :, -6], axis=-1)
# (gt - anchor) / size(anchor) / variance for all four coordinates,
# where 'size' refers to w and h respectively
y_encoded[:, :, -12:-8] /= y_encoded[:, :, -4:]
if diagnostics:
# Here we'll save the matched anchor boxes (i.e. anchor boxes that were matched to a ground truth box,
# but keeping the anchor box coordinates).
y_matched_anchors = np.copy(y_encoded)
# Keeping the anchor box coordinates means setting the offsets to zero.
y_matched_anchors[:, :, -12:-8] = 0
return y_encoded, y_matched_anchors
else:
return y_encoded
def decode_detections_debug(y_pred,
confidence_thresh=0.01,
iou_threshold=0.45,
top_k=200,
input_coords='centroids',
normalize_coords=True,
img_height=None,
img_width=None,
variance_encoded_in_target=False,
border_pixels='half'):
"""
This decoder performs the same processing as `decode_detections()`, but the output format for each left-over
predicted box is `[box_id, class_id, confidence, xmin, ymin, xmax, ymax]`.
That is, in addition to the usual data, each predicted box has the internal index of that box within
the model (`box_id`) prepended to it. This allows you to know exactly which part of the model made a given
box prediction; in particular, it allows you to know which predictor layer made a given prediction.
This can be useful for debugging.
Arguments:
y_pred (array): The prediction output of the SSD model, expected to be a Numpy array
of shape `(batch_size, #boxes, #classes + 4 + 4 + 4)`, where `#boxes` is the total number of
boxes predicted by the model per image and the last axis contains
`[one-hot vector for the classes, 4 predicted coordinate offsets, 4 anchor box coordinates, 4 variances]`.
confidence_thresh (float, optional): A float in [0,1), the minimum classification confidence in a specific
positive class in order to be considered for the non-maximum suppression stage for the respective class.
A lower value will result in a larger part of the selection process being done by the non-maximum
suppression stage, while a larger value will result in a larger part of the selection process happening
in the confidence thresholding stage.
iou_threshold (float, optional): A float in [0,1]. All boxes with a Jaccard similarity of greater than
`iou_threshold` with a locally maximal box will be removed from the set of predictions for a given class,
where 'maximal' refers to the box score.
top_k (int, optional): The number of highest scoring predictions to be kept for each batch item after the
non-maximum suppression stage.
input_coords (str, optional): The box coordinate format that the model outputs. Can be either 'centroids'
for the format `(cx, cy, w, h)` (box center coordinates, width, and height), 'minmax' for the format
`(xmin, xmax, ymin, ymax)`, or 'corners' for the format `(xmin, ymin, xmax, ymax)`.
normalize_coords (bool, optional): Set to `True` if the model outputs relative coordinates
(i.e. coordinates in [0,1]) and you wish to transform these relative coordinates back to absolute
coordinates. If the model outputs relative coordinates, but you do not want to convert them back
to absolute coordinates, set this to `False`. Do not set this to `True` if the model already outputs
absolute coordinates, as that would result in incorrect coordinates. Requires `img_height` and `img_width`
if set to `True`.
img_height (int, optional): The height of the input images. Only needed if `normalize_coords` is `True`.
img_width (int, optional): The width of the input images. Only needed if `normalize_coords` is `True`.
border_pixels (str, optional): How to treat the border pixels of the bounding boxes.
Can be 'include', 'exclude', or 'half'. If 'include', the border pixels belong
to the boxes. If 'exclude', the border pixels do not belong to the boxes.
If 'half', then one of each of the two horizontal and vertical borders belong
to the boxex, but not the other.
Returns:
A python list of length `batch_size` where each list element represents the predicted boxes
for one image and contains a Numpy array of shape `(boxes, 7)` where each row is a box prediction for
a non-background class for the respective image in the format
`[box_id, class_id, confidence, xmin, ymin, xmax, ymax]`.
"""
if normalize_coords and ((img_height is None) or (img_width is None)):
raise ValueError(
"If relative box coordinates are supposed to be converted to absolute coordinates, "
"the decoder needs the image size in order to decode the predictions, "
"but `img_height == {}` and `img_width == {}`".format(
img_height, img_width))
# 1: Convert the box coordinates from the predicted anchor box offsets to predicted absolute coordinates
# Slice out the classes and the four offsets, throw away the anchor coordinates and variances,
# resulting in a tensor of shape `[batch, n_boxes, n_classes + 4 coordinates]`
y_pred_decoded_raw = np.copy(y_pred[:, :, :-8])
if input_coords == 'centroids':
if variance_encoded_in_target:
# Decode the predicted box center x and y coordinates.
y_pred_decoded_raw[:, :, [-4, -3]] = \
y_pred_decoded_raw[:, :, [-4, -3]] * y_pred[:, :, [-6, -5]] + y_pred[:, :, [-8, -7]]
# Decode the predicted box width and heigt.
y_pred_decoded_raw[:, :, [-2, -1]] = \
np.exp(y_pred_decoded_raw[:, :, [-2, -1]]) * y_pred[:, :, [-6, -5]]
else:
# Decode the predicted box center x and y coordinates.
y_pred_decoded_raw[:, :, [-4, -3]] = \
y_pred_decoded_raw[:, :, [-4, -3]] * \
y_pred[:, :, [-6, -5]] * y_pred[:, :, [-4, -3]] + y_pred[:, :, [-8, -7]]
# Decode the predicted box width and heigt.
y_pred_decoded_raw[:, :, [-2, -1]] = np.exp(
y_pred_decoded_raw[:, :, [-2, -1]] *
y_pred[:, :, [-2, -1]]) * y_pred[:, :, [-6, -5]]
y_pred_decoded_raw = convert_coordinates(
y_pred_decoded_raw, start_index=-4, conversion='centroids2corners')
elif input_coords == 'minmax':
# delta(pred) / size(anchor) / variance * variance == delta(pred) / size(anchor) for all four coordinates,
# where 'size' refers to w or h, respectively
y_pred_decoded_raw[:, :, -4:] *= y_pred[:, :, -4:]
# delta_xmin(pred) / w(anchor) * w(anchor) == delta_xmin(pred),
# delta_xmax(pred) / w(anchor) * w(anchor) == delta_xmax(pred)
y_pred_decoded_raw[:, :, [-4, -3]] *= np.expand_dims(y_pred[:, :, -7] -
y_pred[:, :, -8],
axis=-1)
# delta_ymin(pred) / h(anchor) * h(anchor) == delta_ymin(pred),
# delta_ymax(pred) / h(anchor) * h(anchor) == delta_ymax(pred)
y_pred_decoded_raw[:, :, [-2, -1]] *= np.expand_dims(y_pred[:, :, -5] -
y_pred[:, :, -6],
axis=-1)
y_pred_decoded_raw[:, :,
-4:] += y_pred[:, :, -8:
-4] # delta(pred) + anchor == pred for all four coordinates
y_pred_decoded_raw = convert_coordinates(y_pred_decoded_raw,
start_index=-4,
conversion='minmax2corners')
elif input_coords == 'corners':
# delta(pred) / size(anchor) / variance * variance == delta(pred) / size(anchor) for all four coordinates,
# where 'size' refers to w or h, respectively
y_pred_decoded_raw[:, :, -4:] *= y_pred[:, :, -4:]
# delta_xmin(pred) / w(anchor) * w(anchor) == delta_xmin(pred),
# delta_xmax(pred) / w(anchor) * w(anchor) == delta_xmax(pred)
y_pred_decoded_raw[:, :, [-4, -2]] *= np.expand_dims(y_pred[:, :, -6] -
y_pred[:, :, -8],
axis=-1)
# delta_ymin(pred) / h(anchor) * h(anchor) == delta_ymin(pred),
# delta_ymax(pred) / h(anchor) * h(anchor) == delta_ymax(pred)
y_pred_decoded_raw[:, :, [-3, -1]] *= np.expand_dims(y_pred[:, :, -5] -
y_pred[:, :, -7],
axis=-1)
y_pred_decoded_raw[:, :,
-4:] += y_pred[:, :, -8:
-4] # delta(pred) + anchor == pred for all four coordinates
else:
raise ValueError("Unexpected value for `input_coords`. \
Supported input coordinate formats are 'minmax', 'corners' and 'centroids'."
)
# 2: If the model predicts normalized box coordinates and they are supposed to be converted
# back to absolute coordinates, do that
if normalize_coords:
y_pred_decoded_raw[:, :, [
-4, -2
]] *= img_width # Convert xmin, xmax back to absolute coordinates
y_pred_decoded_raw[:, :, [
-3, -1
]] *= img_height # Convert ymin, ymax back to absolute coordinates
# 3: For each batch item, prepend each box's internal index to its coordinates.
y_pred_decoded_raw2 = np.zeros(
(y_pred_decoded_raw.shape[0], y_pred_decoded_raw.shape[1],
y_pred_decoded_raw.shape[2] + 1)) # Expand the last axis by one.
y_pred_decoded_raw2[:, :, 1:] = y_pred_decoded_raw
y_pred_decoded_raw2[:, :, 0] = np.arange(
y_pred_decoded_raw.shape[1]
) # Put the box indices as the first element for each box via broadcasting.
y_pred_decoded_raw = y_pred_decoded_raw2
# 4: Apply confidence thresholding and non-maximum suppression per class
# The number of classes is the length of the last axis minus the four box coordinates and minus the index
n_classes = y_pred_decoded_raw.shape[-1] - 5
y_pred_decoded = [] # Store the final predictions in this list
for batch_item in y_pred_decoded_raw: # `batch_item` has shape `[n_boxes, n_classes + 4 coords]`
pred = [] # Store the final predictions for this batch item here
for class_id in range(
1, n_classes
): # For each class except the background class (which has class ID 0)...
# ...keep only the confidences for that class, making this an array of shape `[n_boxes, 6]` and...
single_class = batch_item[:, [0, class_id + 1, -4, -3, -2, -1]]
# ...keep only those boxes with a confidence above the set threshold.
threshold_met = single_class[single_class[:,
1] > confidence_thresh]
if threshold_met.shape[0] > 0: # If any boxes made the threshold...
maxima = _greedy_nms_debug(
threshold_met,
iou_threshold=iou_threshold,
coords='corners',
border_pixels=border_pixels) # ...perform NMS on them.
# Expand the last dimension by one element to have room for the class ID.
# This is now an arrray of shape `[n_boxes, 6]`
maxima_output = np.zeros(
(maxima.shape[0], maxima.shape[1] + 1))
maxima_output[:,
0] = maxima[:,
0] # Write the box index to the first column...
maxima_output[:,
1] = class_id # ...and write the class ID to the second column...
maxima_output[:,
2:] = maxima[:,
1:] # ...and write the rest of the maxima data to the other columns...
pred.append(
maxima_output
) # ...and append the maxima for this class to the list of maxima for this batch item.
# Once we're through with all classes, keep only the `top_k` maxima with the highest scores
pred = np.concatenate(pred, axis=0)
# If we have more than `top_k` results left at this point, otherwise there is nothing to filter,...
if pred.shape[0] > top_k:
top_k_indices = np.argpartition(
pred[:, 2], kth=pred.shape[0] - top_k, axis=0
)[pred.shape[0] -
top_k:] # ...get the indices of the `top_k` highest-score maxima...
pred = pred[
top_k_indices] # ...and keep only those entries of `pred`...
# and now that we're done, append the array of final predictions for this batch item to the output list
y_pred_decoded.append(pred)
return y_pred_decoded
def decode_detections_fast(y_pred,
confidence_thresh=0.5,
iou_threshold=0.45,
top_k='all',
input_coords='centroids',
normalize_coords=True,
img_height=None,
img_width=None,
border_pixels='half'):
"""
Convert model prediction output back to a format that contains only the positive box predictions
(i.e. the same format that `enconde_y()` takes as input).
Optionally performs confidence thresholding and greedy non-maximum suppression after the decoding stage.
Note that the decoding procedure used here is not the same as the procedure used in the original Caffe
implementation. For each box, the procedure used here assigns the box's highest confidence as its predicted class.
Then it removes all boxes for which the highest confidence is the background class. This results in less work for
the subsequent non-maximum suppression, because the vast majority of the predictions will be filtered out just by
the fact that their highest confidence is for the background class. It is much more efficient than the procedure
of the original implementation, but the results may also differ.
Arguments:
y_pred (array): The prediction output of the SSD model, expected to be a Numpy array
of shape `(batch_size, #boxes, #classes + 4 + 4 + 4)`, where `#boxes` is the total number of
boxes predicted by the model per image and the last axis contains
`[one-hot vector for the classes, 4 predicted coordinate offsets, 4 anchor box coordinates, 4 variances]`.
confidence_thresh (float, optional): A float in [0,1), the minimum classification confidence in any positive
class required for a given box to be considered a positive prediction. A lower value will result
in better recall, while a higher value will result in better precision. Do not use this parameter with the
goal to combat the inevitably many duplicates that an SSD will produce, the subsequent non-maximum
suppression stage will take care of those.
iou_threshold (float, optional): `None` or a float in [0,1]. If `None`, no non-maximum suppression will be
performed. If not `None`, greedy NMS will be performed after the confidence thresholding stage, meaning
all boxes with a Jaccard similarity of greater than `iou_threshold` with a locally maximal box will be
removed from the set of predictions, where 'maximal' refers to the box score.
top_k (int, optional): 'all' or an integer with number of highest scoring predictions to be kept for each
batch item after the non-maximum suppression stage. If 'all', all predictions left after the NMS stage
will be kept.
input_coords (str, optional): The box coordinate format that the model outputs. Can be either 'centroids'
for the format `(cx, cy, w, h)` (box center coordinates, width, and height), 'minmax' for the format
`(xmin, xmax, ymin, ymax)`, or 'corners' for the format `(xmin, ymin, xmax, ymax)`.
normalize_coords (bool, optional): Set to `True` if the model outputs relative coordinates
(i.e. coordinates in [0,1]) and you wish to transform these relative coordinates back to absolute
coordinates. If the model outputs relative coordinates, but you do not want to convert them back
to absolute coordinates, set this to `False`. Do not set this to `True` if the model already outputs
absolute coordinates, as that would result in incorrect coordinates. Requires `img_height` and `img_width`
if set to `True`.
img_height (int, optional): The height of the input images. Only needed if `normalize_coords` is `True`.
img_width (int, optional): The width of the input images. Only needed if `normalize_coords` is `True`.
border_pixels (str, optional): How to treat the border pixels of the bounding boxes.
Can be 'include', 'exclude', or 'half'. If 'include', the border pixels belong
to the boxes. If 'exclude', the border pixels do not belong to the boxes.
If 'half', then one of each of the two horizontal and vertical borders belong
to the boxex, but not the other.
Returns:
A python list of length `batch_size` where each list element represents the predicted boxes
for one image and contains a Numpy array of shape `(boxes, 6)` where each row is a box prediction for
a non-background class for the respective image in the format `[class_id, confidence, xmin, xmax, ymin, ymax]`.
"""
if normalize_coords and ((img_height is None) or (img_width is None)):
raise ValueError(
"If relative box coordinates are supposed to be converted to absolute coordinates, \
the decoder needs the image size in order to decode the predictions, \
but `img_height == {}` and `img_width == {}`".format(
img_height, img_width))
# 1: Convert the classes from one-hot encoding to their class ID
# Slice out the four offset predictions plus two elements whereto
# we'll write the class IDs and confidences in the next step
y_pred_converted = np.copy(y_pred[:, :, -14:-8])
# The indices of the highest confidence values in the one-hot class vectors are the class ID
y_pred_converted[:, :, 0] = np.argmax(y_pred[:, :, :-12], axis=-1)
y_pred_converted[:, :, 1] = np.amax(
y_pred[:, :, :-12],
axis=-1) # Store the confidence values themselves, too
# 2: Convert the box coordinates from the predicted anchor box offsets to predicted absolute coordinates
if input_coords == 'centroids':
# exp(ln(w(pred)/w(anchor)) / w_variance * w_variance) == w(pred) / w(anchor),
# exp(ln(h(pred)/h(anchor)) / h_variance * h_variance) == h(pred) / h(anchor)
y_pred_converted[:, :, [4, 5]] = np.exp(
y_pred_converted[:, :, [4, 5]] * y_pred[:, :, [-2, -1]])
# (w(pred) / w(anchor)) * w(anchor) == w(pred), (h(pred) / h(anchor)) * h(anchor) == h(pred)
y_pred_converted[:, :, [4, 5]] *= y_pred[:, :, [-6, -5]]
# (delta_cx(pred) / w(anchor) / cx_variance) * cx_variance * w(anchor) == delta_cx(pred),
# (delta_cy(pred) / h(anchor) / cy_variance) * cy_variance * h(anchor) == delta_cy(pred)
y_pred_converted[:, :,
[2, 3]] *= y_pred[:, :, [-4, -3]] * y_pred[:, :,
[-6, -5]]
# delta_cx(pred) + cx(anchor) == cx(pred), delta_cy(pred) + cy(anchor) == cy(pred)
y_pred_converted[:, :, [2, 3]] += y_pred[:, :, [-8, -7]]
y_pred_converted = convert_coordinates(y_pred_converted,
start_index=-4,
conversion='centroids2corners')
elif input_coords == 'minmax':
# delta(pred) / size(anchor) / variance * variance == delta(pred) / size(anchor) for all four coordinates,
# where 'size' refers to w or h, respectively
y_pred_converted[:, :, 2:] *= y_pred[:, :, -4:]
# delta_xmin(pred) / w(anchor) * w(anchor) == delta_xmin(pred),
# delta_xmax(pred) / w(anchor) * w(anchor) == delta_xmax(pred)
y_pred_converted[:, :, [2, 3]] *= np.expand_dims(y_pred[:, :, -7] -
y_pred[:, :, -8],
axis=-1)
# delta_ymin(pred) / h(anchor) * h(anchor) == delta_ymin(pred),
# delta_ymax(pred) / h(anchor) * h(anchor) == delta_ymax(pred)
y_pred_converted[:, :, [4, 5]] *= np.expand_dims(y_pred[:, :, -5] -
y_pred[:, :, -6],
axis=-1)
y_pred_converted[:, :,
2:] += y_pred[:, :, -8:
-4] # delta(pred) + anchor == pred for all four coordinates
y_pred_converted = convert_coordinates(y_pred_converted,
start_index=-4,
conversion='minmax2corners')
elif input_coords == 'corners':
# delta(pred) / size(anchor) / variance * variance == delta(pred) / size(anchor) for all four coordinates,
# where 'size' refers to w or h, respectively
y_pred_converted[:, :, 2:] *= y_pred[:, :, -4:]
# delta_xmin(pred) / w(anchor) * w(anchor) == delta_xmin(pred),
# delta_xmax(pred) / w(anchor) * w(anchor) == delta_xmax(pred)
y_pred_converted[:, :, [2, 4]] *= np.expand_dims(y_pred[:, :, -6] -
y_pred[:, :, -8],
axis=-1)
# delta_ymin(pred) / h(anchor) * h(anchor) == delta_ymin(pred),
# delta_ymax(pred) / h(anchor) * h(anchor) == delta_ymax(pred)
y_pred_converted[:, :, [3, 5]] *= np.expand_dims(y_pred[:, :, -5] -
y_pred[:, :, -7],
axis=-1)
y_pred_converted[:, :,
2:] += y_pred[:, :, -8:
-4] # delta(pred) + anchor == pred for all four coordinates
else:
raise ValueError(
"Unexpected value for `coords`. Supported values are 'minmax', 'corners' and 'centroids'."
)
# 3: If the model predicts normalized box coordinates and
# they are supposed to be converted back to absolute coordinates, do that
if normalize_coords:
y_pred_converted[:, :, [
2, 4
]] *= img_width # Convert xmin, xmax back to absolute coordinates
y_pred_converted[:, :, [
3, 5
]] *= img_height # Convert ymin, ymax back to absolute coordinates
# 4: Decode our huge `(batch, #boxes, 6)` tensor into a list of length
# `batch` where each list entry is an array containing only the positive predictions
y_pred_decoded = []
for batch_item in y_pred_converted: # For each image in the batch...
# ...get all boxes that don't belong to the background class,...
boxes = batch_item[np.nonzero(batch_item[:, 0])]
# ...then filter out those positive boxes for which the prediction confidence is too low and after that...
boxes = boxes[boxes[:, 1] >= confidence_thresh]
if iou_threshold: # ...if an IoU threshold is set...
boxes = _greedy_nms2(boxes,
iou_threshold=iou_threshold,
coords='corners',
border_pixels=border_pixels
) # ...perform NMS on the remaining boxes.
if top_k != 'all' and boxes.shape[
0] > top_k: # If we have more than `top_k` results left at this point...
top_k_indices = np.argpartition(
boxes[:, 1], kth=boxes.shape[0] - top_k, axis=0
)[boxes.shape[0] -
top_k:] # ...get the indices of the `top_k` highest-scoring boxes...
boxes = boxes[top_k_indices] # ...and keep only those boxes...
# ...and now that we're done, append the array of final predictions for this batch item to the output list
y_pred_decoded.append(boxes)
return y_pred_decoded
def decode_detections(y_pred,
confidence_thresh=0.01,
iou_threshold=0.45,
top_k=200,
input_coords='centroids',
normalize_coords=True,
img_height=None,
img_width=None,
border_pixels='half'):
"""
Convert model prediction output back to a format that contains only the positive box predictions
(i.e. the same format that `SSDInputEncoder` takes as input).
After the decoding, two stages of prediction filtering are performed for each class individually:
First confidence thresholding, then greedy non-maximum suppression. The filtering results for all
classes are concatenated and the `top_k` overall highest confidence results constitute the final
predictions for a given batch item. This procedure follows the original Caffe implementation.
For a slightly different and more efficient alternative to decode raw model output that performs
non-maximum suppresion globally instead of per class, see `decode_detections_fast()` below.
Arguments:
y_pred (array): The prediction output of the SSD model, expected to be a Numpy array
of shape `(batch_size, #boxes, #classes + 4 + 4 + 4)`, where `#boxes` is the total number of
boxes predicted by the model per image and the last axis contains
`[one-hot vector for the classes, 4 predicted coordinate offsets, 4 anchor box coordinates, 4 variances]`.
confidence_thresh (float, optional): A float in [0,1), the minimum classification confidence in a specific
positive class in order to be considered for the non-maximum suppression stage for the respective class.
A lower value will result in a larger part of the selection process being done by the non-maximum
suppression stage, while a larger value will result in a larger part of the selection process happening
in the confidence thresholding stage.
iou_threshold (float, optional): A float in [0,1]. All boxes with a Jaccard similarity of greater than
`iou_threshold` with a locally maximal box will be removed from the set of predictions for a given class,
where 'maximal' refers to the box score.
top_k (int, optional): The number of highest scoring predictions to be kept for each batch item after the
non-maximum suppression stage.
input_coords (str, optional): The box coordinate format that the model outputs. Can be either 'centroids'
for the format `(cx, cy, w, h)` (box center coordinates, width, and height), 'minmax' for the format
`(xmin, xmax, ymin, ymax)`, or 'corners' for the format `(xmin, ymin, xmax, ymax)`.
normalize_coords (bool, optional): Set to `True` if the model outputs relative coordinates
(i.e. coordinates in [0,1]) and you wish to transform these relative coordinates back to absolute
coordinates. If the model outputs relative coordinates, but you do not want to convert them back to
absolute coordinates, set this to `False`. Do not set this to `True` if the model already outputs
absolute coordinates, as that would result in incorrect coordinates. Requires `img_height`
and `img_width` if set to `True`.
img_height (int, optional): The height of the input images. Only needed if `normalize_coords` is `True`.
img_width (int, optional): The width of the input images. Only needed if `normalize_coords` is `True`.
border_pixels (str, optional): How to treat the border pixels of the bounding boxes.
Can be 'include', 'exclude', or 'half'. If 'include', the border pixels belong
to the boxes. If 'exclude', the border pixels do not belong to the boxes.
If 'half', then one of each of the two horizontal and vertical borders belong
to the boxex, but not the other.
Returns:
A python list of length `batch_size` where each list element represents the predicted boxes
for one image and contains a Numpy array of shape `(boxes, 6)` where each row is a box prediction for
a non-background class for the respective image in the format `[class_id, confidence, xmin, ymin, xmax, ymax]`.
"""
if normalize_coords and ((img_height is None) or (img_width is None)):
raise ValueError(
"If relative box coordinates are supposed to be converted to absolute coordinates, \
the decoder needs the image size in order to decode the predictions, \
but `img_height == {}` and `img_width == {}`".format(
img_height, img_width))
# 1: Convert the box coordinates from the predicted anchor box offsets to predicted absolute coordinates
# Slice out the classes and the four offsets, throw away the anchor coordinates and variances,
# resulting in a tensor of shape `[batch, n_boxes, n_classes + 4 coordinates]`
y_pred_decoded_raw = np.copy(y_pred[:, :, :-8])
if input_coords == 'centroids':
# exp(ln(w(pred)/w(anchor)) / w_variance * w_variance) ==
# w(pred) / w(anchor), exp(ln(h(pred)/h(anchor)) / h_variance * h_variance) == h(pred) / h(anchor)
y_pred_decoded_raw[:, :, [-2, -1]] = np.exp(
y_pred_decoded_raw[:, :, [-2, -1]] * y_pred[:, :, [-2, -1]])
# (w(pred) / w(anchor)) * w(anchor) == w(pred), (h(pred) / h(anchor)) * h(anchor) == h(pred)
y_pred_decoded_raw[:, :, [-2, -1]] *= y_pred[:, :, [-6, -5]]
# (delta_cx(pred) / w(anchor) / cx_variance) * cx_variance * w(anchor) == delta_cx(pred),
# (delta_cy(pred) / h(anchor) / cy_variance) * cy_variance * h(anchor) == delta_cy(pred)
y_pred_decoded_raw[:, :,
[-4, -3]] *= y_pred[:, :,
[-4, -3]] * y_pred[:, :,
[-6, -5]]
# delta_cx(pred) + cx(anchor) == cx(pred), delta_cy(pred) + cy(anchor) == cy(pred)
y_pred_decoded_raw[:, :, [-4, -3]] += y_pred[:, :, [-8, -7]]
y_pred_decoded_raw = convert_coordinates(
y_pred_decoded_raw, start_index=-4, conversion='centroids2corners')
elif input_coords == 'minmax':
# delta(pred) / size(anchor) / variance * variance == delta(pred) / size(anchor) for all four coordinates,
# where 'size' refers to w or h, respectively
y_pred_decoded_raw[:, :, -4:] *= y_pred[:, :, -4:]
# delta_xmin(pred) / w(anchor) * w(anchor) == delta_xmin(pred),
# delta_xmax(pred) / w(anchor) * w(anchor) == delta_xmax(pred)
y_pred_decoded_raw[:, :, [-4, -3]] *= np.expand_dims(y_pred[:, :, -7] -
y_pred[:, :, -8],
axis=-1)
# delta_ymin(pred) / h(anchor) * h(anchor) == delta_ymin(pred),
# delta_ymax(pred) / h(anchor) * h(anchor) == delta_ymax(pred)
y_pred_decoded_raw[:, :, [-2, -1]] *= np.expand_dims(y_pred[:, :, -5] -
y_pred[:, :, -6],
axis=-1)
y_pred_decoded_raw[:, :,
-4:] += y_pred[:, :, -8:
-4] # delta(pred) + anchor == pred for all four coordinates
y_pred_decoded_raw = convert_coordinates(y_pred_decoded_raw,
start_index=-4,
conversion='minmax2corners')
elif input_coords == 'corners':
# delta(pred) / size(anchor) / variance * variance == delta(pred) / size(anchor) for all four coordinates,
# where 'size' refers to w or h, respectively
y_pred_decoded_raw[:, :, -4:] *= y_pred[:, :, -4:]
# delta_xmin(pred) / w(anchor) * w(anchor) == delta_xmin(pred),
# delta_xmax(pred) / w(anchor) * w(anchor) == delta_xmax(pred)
y_pred_decoded_raw[:, :, [-4, -2]] *= np.expand_dims(y_pred[:, :, -6] -
y_pred[:, :, -8],
axis=-1)
# delta_ymin(pred) / h(anchor) * h(anchor) == delta_ymin(pred),
# delta_ymax(pred) / h(anchor) * h(anchor) == delta_ymax(pred)
y_pred_decoded_raw[:, :, [-3, -1]] *= np.expand_dims(y_pred[:, :, -5] -
y_pred[:, :, -7],
axis=-1)
y_pred_decoded_raw[:, :,
-4:] += y_pred[:, :, -8:
-4] # delta(pred) + anchor == pred for all four coordinates
else:
raise ValueError("Unexpected value for `input_coords`. \
Supported input coordinate formats are 'minmax', 'corners' and 'centroids'."
)
# 2: If the model predicts normalized box coordinates and they are supposed
# to be converted back to absolute coordinates, do that
if normalize_coords:
y_pred_decoded_raw[:, :, [
-4, -2
]] *= img_width # Convert xmin, xmax back to absolute coordinates
y_pred_decoded_raw[:, :, [
-3, -1
]] *= img_height # Convert ymin, ymax back to absolute coordinates
# 3: Apply confidence thresholding and non-maximum suppression per class
n_classes = y_pred_decoded_raw.shape[
-1] - 4 # The number of classes is the length of the last axis minus the four box coordinates
y_pred_decoded = [] # Store the final predictions in this list
for batch_item in y_pred_decoded_raw: # `batch_item` has shape `[n_boxes, n_classes + 4 coords]`
pred = [] # Store the final predictions for this batch item here
for class_id in range(
1, n_classes
): # For each class except the background class (which has class ID 0)...
# ...keep only the confidences for that class, making this an array of shape `[n_boxes, 5]` and...
single_class = batch_item[:, [class_id, -4, -3, -2, -1]]
# ...keep only those boxes with a confidence above the set threshold.
threshold_met = single_class[single_class[:,
0] > confidence_thresh]
if threshold_met.shape[0] > 0: # If any boxes made the threshold...
maxima = _greedy_nms(
threshold_met,
iou_threshold=iou_threshold,
coords='corners',
border_pixels=border_pixels) # ...perform NMS on them.
# Expand the last dimension by one element to have room for the class ID.
# This is now an arrray of shape `[n_boxes, 6]`
maxima_output = np.zeros(
(maxima.shape[0], maxima.shape[1] + 1))
maxima_output[:,
0] = class_id # Write the class ID to the first column...
maxima_output[:,
1:] = maxima # ...and write the maxima to the other columns...
# ...and append the maxima for this class to the list of maxima for this batch item.
pred.append(maxima_output)
# Once we're through with all classes, keep only the `top_k` maxima with the highest scores
if pred: # If there are any predictions left after confidence-thresholding...
pred = np.concatenate(pred, axis=0)
# If we have more than `top_k` results left at this point, otherwise there is nothing to filter,...
if top_k != 'all' and pred.shape[0] > top_k:
top_k_indices = np.argpartition(
pred[:, 1], kth=pred.shape[0] - top_k, axis=0
)[pred.shape[0] -
top_k:] # ...get the indices of the `top_k` highest-score maxima...
pred = pred[
top_k_indices] # ...and keep only those entries of `pred`...
else:
pred = np.array(
pred) # Even if empty, `pred` must become a Numpy array.
# ...and now that we're done, append the array of final predictions for this batch item to the output list
y_pred_decoded.append(pred)
return y_pred_decoded
def call(self, x, mask=None):
"""
Return an anchor box based on the shape of the input tensor.
:param x: 4D tensor of shape (batch, channels, height, width) if dim_ordering = "th"
or (batch, height, width, channels) if dim_ordering = "tf". The input for this
layer must be the output of the localization predictor layer.
:param mask:
:return:
"""
# Compute box width and height for each aspect ratio.
# The shorter side of the image will be used to compute 'w' and 'h'
# using 'scale' and 'aspect_ratios'.
size = min(self.img_height, self.img_width)
# Compute the box widths and heights for all aspect ratios
wh_list = []
for ar in self.aspect_ratios:
if ar == 1:
# Compute the regular anchor box for aspect ratio 1
box_height = box_width = self.this_scale * size
wh_list.append((box_width, box_height))
if self.two_boxes_for_ar1:
# Compute one slightly larger version using the geometric mean of
# this scale and the next
box_height = box_width = np.sqrt(self.this_scale * self.next_scale) * size
wh_list.append((box_width, box_height))
else:
box_height = self.this_scale * size * np.sqrt(ar)
box_width = self.this_scale * size / np.sqrt(ar)
wh_list.append((box_width, box_height))
wh_list = np.array(wh_list)
# Get the shape of the input tensor
if K.image_data_format() == 'channels_last': # notice this, different version, different condition
batch_size, feature_map_height, feature_map_width, feature_map_channels = x._keras_shape
else:
batch_size, feature_map_channels, feature_map_height, feature_map_width = x._keras_shape
# Compute the grid of box center points.
# They are identical to all aspect ratios.
# Compute the step size, i.e. how far apart the anchor box center points will be vertically and horizontally
if self.this_steps is None:
step_height = self.img_height / feature_map_height
step_width = self.img_width / feature_map_width
else:
if isinstance(self.this_steps, (list, tuple)) and len(self.this_steps) == 2:
step_height = self.this_steps[0]
step_width = self.this_steps[1]
elif isinstance(self.this_steps, (int, float)):
step_height = self.this_steps
step_width = self.this_steps
# Compute the offsets, i.e. at what pixel values the first anchor box center point
# will be from the top and from the left of the image.
if self.this_offsets is None:
offset_height = 0.5
offset_width = 0.5
else:
if isinstance(self.this_offsets, (list, tuple)) and len(self.this_offsets) == 2:
offset_height = self.this_offsets[0]
offset_width = self.this_offsets[1]
elif isinstance(self.this_offsets, (int, float)):
offset_height = self.this_offsets
offset_width = self.this_offsets
# Now that we have the step sizes and offsets,
# compute the grid of anchor box center points.
cy = np.linspace(offset_height * step_height,
(offset_height + feature_map_height - 1) * step_height,
feature_map_height)
cx = np.linspace(offset_width * step_width,
(offset_width + feature_map_width - 1) * step_width,
feature_map_width)
cx_grid, cy_grid = np.meshgrid(cx, cy)
# This is necessary for np.tile() to do what we want future down
cx_grid = np.expand_dims(cx_grid, -1)
cy_grid = np.expand_dims(cy_grid, -1)
# Create a 4D tensor template of the shape
# (feature_map_height, feature_map_width, n_boxes, 4)
# where the last dimension will contain (cx, cy, w, h)
boxes_tensor = np.zeros((feature_map_height, feature_map_width, self.n_boxes, 4))
boxes_tensor[:, :, :, 0] = np.tile(cx_grid, (1, 1, self.n_boxes)) # Set cx
boxes_tensor[:, :, :, 1] = np.tile(cy_grid, (1, 1, self.n_boxes)) # Set cy
boxes_tensor[:, :, :, 2] = wh_list[:, 0] # Set w
boxes_tensor[:, :, :, 3] = wh_list[:, 1] # Set h
# Convert (cx, cy, w, h) to (xmin, xmax, ymin, ymax)
boxes_tensor = convert_coordinates(boxes_tensor, start_index=0, conversion='centroids2corners')
# If clip_boxes is enabled, clip the coordinates to lie within the image boundaries
if self.clip_boxes:
x_coords = boxes_tensor[:, :, :, [0, 2]]
x_coords[x_coords >= self.img_width] = self.img_width - 1
x_coords[x_coords < 0] = 0
boxes_tensor[:, :, :, [0, 2]] = x_coords
y_coords = boxes_tensor[:, :, :, [1, 3]]
y_coords[y_coords >= self.img_height] = self.img_height - 1
y_coords[y_coords < 0] = 0
boxes_tensor[:, :, :, [1, 3]] = y_coords
# If normalize_coords is enabled, normalize the coordinates to be within [0,1]
if self.normalize_coords:
boxes_tensor[:, :, :, [0, 2]] /= self.img_width
boxes_tensor[:, :, :, [1, 3]] /= self.img_height
if self.coords == 'centroids':
# Convert (xmin, ymin, xmax, ymax) back to (cx, cy, w, h)
boxes_tensor = convert_coordinates(boxes_tensor, start_index=0, conversion='corners2centroids')
elif self.coords == 'minmax':
# Convert (xmin, ymin, xmax, ymax) to (xmin, ymin, xmax, ymax)
boxes_tensor = convert_coordinates(boxes_tensor, start_index=0, conversion='corners2minmax')
# Create a tensor to contain the variances and append it to boxes_tensor.
# This tensor has the same shape as boxes_tensor and simply contains the same
# 4 variance values for every position in the last axis.
variances_tensor = np.zeros_like(boxes_tensor)
variances_tensor += self.variances
# Concatenate into the shape of (feature_map_height, feature_map_width, n_boxes, 8).
boxes_tensor = np.concatenate((boxes_tensor, variances_tensor), axis=-1)
# Prepend one dimension to boxes_tensor to account for the batch size and tile it along.
# The result will be a 5D tensor of shape (batch_size, feature_map_height, feature_map_width, n_boxes, 8).
boxes_tensor = np.expand_dims(boxes_tensor, axis=0)
boxes_tensor = K.tile(K.constant(boxes_tensor, dtype='float32'), (K.shape(x)[0], 1, 1, 1, 1))
return boxes_tensor