def call(self, inputs):
box_outputs, class_outputs, anchor_boxes, image_shape = inputs
# Collects outputs from all levels into a list.
boxes = []
scores = []
for i in range(self._min_level, self._max_level + 1):
batch_size = tf.shape(input=class_outputs[i])[0]
# Applies score transformation and remove the implicit background class.
scores_i = _apply_score_activation(
class_outputs[i], self._num_classes, self._score_activation)
# Box decoding.
# The anchor boxes are shared for all data in a batch.
# One stage detector only supports class agnostic box regression.
anchor_boxes_i = tf.reshape(anchor_boxes[i], [batch_size, -1, 4])
box_outputs_i = tf.reshape(box_outputs[i], [batch_size, -1, 4])
boxes_i = box_utils.decode_boxes(box_outputs_i, anchor_boxes_i)
# Box clipping.
boxes_i = box_utils.clip_boxes(boxes_i, image_shape)
boxes.append(boxes_i)
scores.append(scores_i)
boxes = tf.concat(boxes, axis=1)
scores = tf.concat(scores, axis=1)
boxes = tf.expand_dims(boxes, axis=2)
(nmsed_boxes, nmsed_scores, nmsed_classes,
valid_detections) = self._generate_detections(
tf.cast(boxes, tf.float32), tf.cast(scores, tf.float32))
# Adds 1 to offset the background class which has index 0.
nmsed_classes += 1
return nmsed_boxes, nmsed_scores, nmsed_classes, valid_detections
def call(self, box_outputs, class_outputs, anchor_boxes, image_shape):
# Collects outputs from all levels into a list.
boxes = []
scores = []
for i in range(self._min_level, self._max_level + 1):
box_outputs_i_shape = tf.shape(box_outputs[i])
batch_size = box_outputs_i_shape[0]
num_anchors_per_locations = box_outputs_i_shape[-1] // 4
num_classes = tf.shape(
class_outputs[i])[-1] // num_anchors_per_locations
# Applies score transformation and remove the implicit background class.
scores_i = tf.sigmoid(
tf.reshape(class_outputs[i], [batch_size, -1, num_classes]))
scores_i = tf.slice(scores_i, [0, 0, 1], [-1, -1, -1])
# Box decoding.
# The anchor boxes are shared for all data in a batch.
# One stage detector only supports class agnostic box regression.
anchor_boxes_i = tf.reshape(anchor_boxes[i], [batch_size, -1, 4])
box_outputs_i = tf.reshape(box_outputs[i], [batch_size, -1, 4])
boxes_i = box_utils.decode_boxes(box_outputs_i, anchor_boxes_i)
# Box clipping.
boxes_i = box_utils.clip_boxes(boxes_i, image_shape)
boxes.append(boxes_i)
scores.append(scores_i)
boxes = tf.concat(boxes, axis=1)
scores = tf.concat(scores, axis=1)
nmsed_boxes, nmsed_scores, nmsed_classes, valid_detections = (
self._generate_detections(tf.expand_dims(boxes, axis=2), scores))
# Adds 1 to offset the background class which has index 0.
nmsed_classes += 1
return nmsed_boxes, nmsed_scores, nmsed_classes, valid_detections
def __call__(self,
box_outputs,
class_outputs,
anchor_boxes,
image_shape,
is_single_fg_score=False,
keep_nms=True):
"""Generate final detections for Object Localization Network (OLN).
Args:
box_outputs: a tensor of shape of [batch_size, K, num_classes * 4]
representing the class-specific box coordinates relative to anchors.
class_outputs: a tensor of shape of [batch_size, K, num_classes]
representing the class logits before applying score activiation.
anchor_boxes: a tensor of shape of [batch_size, K, 4] representing the
corresponding anchor boxes w.r.t `box_outputs`.
image_shape: a tensor of shape of [batch_size, 2] storing the image height
and width w.r.t. the scaled image, i.e. the same image space as
`box_outputs` and `anchor_boxes`.
is_single_fg_score: a Bool indicator of whether class_outputs includes the
background scores concatenated or not. By default, class_outputs is a
concatenation of both scores for the foreground and background. That is,
scores_without_bg=False.
keep_nms: a Bool indicator of whether to perform NMS or not.
Returns:
nms_boxes: `float` Tensor of shape [batch_size, max_total_size, 4]
representing top detected boxes in [y1, x1, y2, x2].
nms_scores: `float` Tensor of shape [batch_size, max_total_size]
representing sorted confidence scores for detected boxes. The values are
between [0, 1].
nms_classes: `int` Tensor of shape [batch_size, max_total_size]
representing classes for detected boxes.
valid_detections: `int` Tensor of shape [batch_size] only the top
`valid_detections` boxes are valid detections.
"""
if is_single_fg_score:
# Concatenates dummy background scores.
dummy_bg_scores = tf.zeros_like(class_outputs)
class_outputs = tf.stack([dummy_bg_scores, class_outputs], -1)
else:
class_outputs = tf.nn.softmax(class_outputs, axis=-1)
# Removes the background class.
class_outputs_shape = tf.shape(class_outputs)
batch_size = class_outputs_shape[0]
num_locations = class_outputs_shape[1]
num_classes = class_outputs_shape[-1]
num_detections = num_locations * (num_classes - 1)
class_outputs = tf.slice(class_outputs, [0, 0, 1], [-1, -1, -1])
box_outputs = tf.reshape(
box_outputs,
tf.stack([batch_size, num_locations, num_classes, 4], axis=-1))
box_outputs = tf.slice(box_outputs, [0, 0, 1, 0], [-1, -1, -1, -1])
anchor_boxes = tf.tile(tf.expand_dims(anchor_boxes, axis=2),
[1, 1, num_classes - 1, 1])
box_outputs = tf.reshape(
box_outputs, tf.stack([batch_size, num_detections, 4], axis=-1))
anchor_boxes = tf.reshape(
anchor_boxes, tf.stack([batch_size, num_detections, 4], axis=-1))
# Box decoding. For RPN outputs, box_outputs are all zeros.
decoded_boxes = box_utils.decode_boxes(box_outputs,
anchor_boxes,
weights=[10.0, 10.0, 5.0, 5.0])
# Box clipping
decoded_boxes = box_utils.clip_boxes(decoded_boxes, image_shape)
decoded_boxes = tf.reshape(
decoded_boxes,
tf.stack([batch_size, num_locations, num_classes - 1, 4], axis=-1))
if keep_nms:
nmsed_boxes, nmsed_scores, nmsed_classes, valid_detections = (
self._generate_detections(decoded_boxes, class_outputs))
# Adds 1 to offset the background class which has index 0.
nmsed_classes += 1
else:
nmsed_boxes = decoded_boxes[:, :, 0, :]
nmsed_scores = class_outputs[:, :, 0]
nmsed_classes = tf.cast(tf.ones_like(nmsed_scores), tf.int32)
valid_detections = tf.cast(
tf.reduce_sum(tf.ones_like(nmsed_scores), axis=-1), tf.int32)
return nmsed_boxes, nmsed_scores, nmsed_classes, valid_detections
def call(self, box_outputs, class_outputs, anchor_boxes, image_shape):
"""Generate final detections.
Args:
box_outputs: a tensor of shape of [batch_size, K, num_classes * 4]
representing the class-specific box coordinates relative to anchors.
class_outputs: a tensor of shape of [batch_size, K, num_classes]
representing the class logits before applying score activiation.
anchor_boxes: a tensor of shape of [batch_size, K, 4] representing the
corresponding anchor boxes w.r.t `box_outputs`.
image_shape: a tensor of shape of [batch_size, 2] storing the image height
and width w.r.t. the scaled image, i.e. the same image space as
`box_outputs` and `anchor_boxes`.
Returns:
nms_boxes: `float` Tensor of shape [batch_size, max_total_size, 4]
representing top detected boxes in [y1, x1, y2, x2].
nms_scores: `float` Tensor of shape [batch_size, max_total_size]
representing sorted confidence scores for detected boxes. The values are
between [0, 1].
nms_classes: `int` Tensor of shape [batch_size, max_total_size]
representing classes for detected boxes.
valid_detections: `int` Tensor of shape [batch_size] only the top
`valid_detections` boxes are valid detections.
"""
class_outputs = tf.nn.softmax(class_outputs, axis=-1)
# Removes the background class.
class_outputs_shape = tf.shape(class_outputs)
batch_size = class_outputs_shape[0]
num_locations = class_outputs_shape[1]
num_classes = class_outputs_shape[-1]
num_detections = num_locations * (num_classes - 1)
class_outputs = tf.slice(class_outputs, [0, 0, 1], [-1, -1, -1])
box_outputs = tf.reshape(
box_outputs,
tf.stack([batch_size, num_locations, num_classes, 4], axis=-1))
box_outputs = tf.slice(box_outputs, [0, 0, 1, 0], [-1, -1, -1, -1])
anchor_boxes = tf.tile(tf.expand_dims(anchor_boxes, axis=2),
[1, 1, num_classes - 1, 1])
box_outputs = tf.reshape(
box_outputs, tf.stack([batch_size, num_detections, 4], axis=-1))
anchor_boxes = tf.reshape(
anchor_boxes, tf.stack([batch_size, num_detections, 4], axis=-1))
# Box decoding.
decoded_boxes = box_utils.decode_boxes(box_outputs,
anchor_boxes,
weights=[10.0, 10.0, 5.0, 5.0])
# Box clipping
decoded_boxes = box_utils.clip_boxes(decoded_boxes, image_shape)
decoded_boxes = tf.reshape(
decoded_boxes,
tf.stack([batch_size, num_locations, num_classes - 1, 4], axis=-1))
nmsed_boxes, nmsed_scores, nmsed_classes, valid_detections = (
self._generate_detections(decoded_boxes, class_outputs))
# Adds 1 to offset the background class which has index 0.
nmsed_classes += 1
return nmsed_boxes, nmsed_scores, nmsed_classes, valid_detections
def oln_multilevel_propose_rois(
self,
rpn_boxes,
rpn_scores,
anchor_boxes,
image_shape,
rpn_pre_nms_top_k=2000,
rpn_post_nms_top_k=1000,
rpn_nms_threshold=0.7,
rpn_score_threshold=0.0,
rpn_min_size_threshold=0.0,
decode_boxes=True,
clip_boxes=True,
use_batched_nms=False,
apply_sigmoid_to_score=True,
is_box_lrtb=False,
rpn_object_scores=None,
):
"""Proposes RoIs given a group of candidates from different FPN levels.
The following describes the steps:
1. For each individual level:
a. Adjust scores for each level if specified by rpn_object_scores.
b. Apply sigmoid transform if specified.
c. Decode boxes (either of xyhw or left-right-top-bottom format) if
specified.
d. Clip boxes if specified.
e. Filter small boxes and those fall outside image if specified.
f. Apply pre-NMS filtering including pre-NMS top k and score
thresholding.
g. Apply NMS.
2. Aggregate post-NMS boxes from each level.
3. Apply an overall top k to generate the final selected RoIs.
Args:
rpn_boxes: a dict with keys representing FPN levels and values
representing box tenors of shape [batch_size, feature_h, feature_w,
num_anchors * 4].
rpn_scores: a dict with keys representing FPN levels and values
representing logit tensors of shape [batch_size, feature_h, feature_w,
num_anchors].
anchor_boxes: a dict with keys representing FPN levels and values
representing anchor box tensors of shape [batch_size, feature_h,
feature_w, num_anchors * 4].
image_shape: a tensor of shape [batch_size, 2] where the last dimension
are [height, width] of the scaled image.
rpn_pre_nms_top_k: an integer of top scoring RPN proposals *per level* to
keep before applying NMS. Default: 2000.
rpn_post_nms_top_k: an integer of top scoring RPN proposals *in total* to
keep after applying NMS. Default: 1000.
rpn_nms_threshold: a float between 0 and 1 representing the IoU threshold
used for NMS. If 0.0, no NMS is applied. Default: 0.7.
rpn_score_threshold: a float between 0 and 1 representing the minimal box
score to keep before applying NMS. This is often used as a pre-filtering
step for better performance. If 0, no filtering is applied. Default: 0.
rpn_min_size_threshold: a float representing the minimal box size in each
side (w.r.t. the scaled image) to keep before applying NMS. This is
often used as a pre-filtering step for better performance. If 0, no
filtering is applied. Default: 0.
decode_boxes: a boolean indicating whether `rpn_boxes` needs to be decoded
using `anchor_boxes`. If False, use `rpn_boxes` directly and ignore
`anchor_boxes`. Default: True.
clip_boxes: a boolean indicating whether boxes are first clipped to the
scaled image size before appliying NMS. If False, no clipping is applied
and `image_shape` is ignored. Default: True.
use_batched_nms: a boolean indicating whether NMS is applied in batch
using `tf.image.combined_non_max_suppression`. Currently only available
in CPU/GPU. Default: False.
apply_sigmoid_to_score: a boolean indicating whether apply sigmoid to
`rpn_scores` before applying NMS. Default: True.
is_box_lrtb: a bool indicating whether boxes are in lrtb (=left,right,top,
bottom) format.
rpn_object_scores: a predicted objectness score (e.g., centerness). In
OLN, we use object_scores=centerness as a replacement of the scores at
each level. A dict with keys representing FPN levels and values
representing logit tensors of shape [batch_size, feature_h, feature_w,
num_anchors].
Returns:
selected_rois: a tensor of shape [batch_size, rpn_post_nms_top_k, 4],
representing the box coordinates of the selected proposals w.r.t. the
scaled image.
selected_roi_scores: a tensor of shape [batch_size, rpn_post_nms_top_k,
1],representing the scores of the selected proposals.
"""
with tf.name_scope('multilevel_propose_rois'):
rois = []
roi_scores = []
image_shape = tf.expand_dims(image_shape, axis=1)
for level in sorted(rpn_scores.keys()):
with tf.name_scope('level_%d' % level):
_, feature_h, feature_w, num_anchors_per_location = (
rpn_scores[level].get_shape().as_list())
num_boxes = feature_h * feature_w * num_anchors_per_location
this_level_scores = tf.reshape(rpn_scores[level],
[-1, num_boxes])
this_level_boxes = tf.reshape(rpn_boxes[level],
[-1, num_boxes, 4])
this_level_anchors = tf.cast(tf.reshape(
anchor_boxes[level], [-1, num_boxes, 4]),
dtype=this_level_scores.dtype)
if rpn_object_scores:
this_level_object_scores = rpn_object_scores[level]
this_level_object_scores = tf.reshape(
this_level_object_scores, [-1, num_boxes])
this_level_object_scores = tf.cast(
this_level_object_scores, this_level_scores.dtype)
this_level_scores = this_level_object_scores
if apply_sigmoid_to_score:
this_level_scores = tf.sigmoid(this_level_scores)
if decode_boxes:
if is_box_lrtb: # Box in left-right-top-bottom format.
this_level_boxes = box_utils.decode_boxes_lrtb(
this_level_boxes, this_level_anchors)
else: # Box in standard x-y-h-w format.
this_level_boxes = box_utils.decode_boxes(
this_level_boxes, this_level_anchors)
if clip_boxes:
this_level_boxes = box_utils.clip_boxes(
this_level_boxes, image_shape)
if rpn_min_size_threshold > 0.0:
this_level_boxes, this_level_scores = box_utils.filter_boxes(
this_level_boxes, this_level_scores, image_shape,
rpn_min_size_threshold)
this_level_pre_nms_top_k = min(num_boxes,
rpn_pre_nms_top_k)
this_level_post_nms_top_k = min(num_boxes,
rpn_post_nms_top_k)
if rpn_nms_threshold > 0.0:
if use_batched_nms:
this_level_rois, this_level_roi_scores, _, _ = (
tf.image.combined_non_max_suppression(
tf.expand_dims(this_level_boxes, axis=2),
tf.expand_dims(this_level_scores, axis=-1),
max_output_size_per_class=
this_level_pre_nms_top_k,
max_total_size=this_level_post_nms_top_k,
iou_threshold=rpn_nms_threshold,
score_threshold=rpn_score_threshold,
pad_per_class=False,
clip_boxes=False))
else:
if rpn_score_threshold > 0.0:
this_level_boxes, this_level_scores = (
box_utils.filter_boxes_by_scores(
this_level_boxes, this_level_scores,
rpn_score_threshold))
this_level_boxes, this_level_scores = box_utils.top_k_boxes(
this_level_boxes,
this_level_scores,
k=this_level_pre_nms_top_k)
this_level_roi_scores, this_level_rois = (
nms.sorted_non_max_suppression_padded(
this_level_scores,
this_level_boxes,
max_output_size=this_level_post_nms_top_k,
iou_threshold=rpn_nms_threshold))
else:
this_level_rois, this_level_roi_scores = box_utils.top_k_boxes(
this_level_rois,
this_level_scores,
k=this_level_post_nms_top_k)
rois.append(this_level_rois)
roi_scores.append(this_level_roi_scores)
all_rois = tf.concat(rois, axis=1)
all_roi_scores = tf.concat(roi_scores, axis=1)
with tf.name_scope('top_k_rois'):
_, num_valid_rois = all_roi_scores.get_shape().as_list()
overall_top_k = min(num_valid_rois, rpn_post_nms_top_k)
selected_rois, selected_roi_scores = box_utils.top_k_boxes(
all_rois, all_roi_scores, k=overall_top_k)
return selected_rois, selected_roi_scores
