def _get_proposals_single(self,
rpn_probs,
rpn_deltas,
mlvl_anchors,
img_shape,
with_probs,
training=True):
"""
Calculate proposals per image
Args:
Returns:
"""
if training:
num_pre_nms = self.num_pre_nms_train
proposal_count = self.num_post_nms_train
else:
num_pre_nms = self.num_pre_nms_test
proposal_count = self.num_post_nms_test
level_ids = []
mlvl_scores = []
mlvl_deltas = []
mlvl_valid_anchors = []
mlvl_proposals = []
num_levels = len(rpn_probs)
for idx in range(num_levels):
level_probs = tf.reshape(rpn_probs[idx],
[-1, 2]) # H, W, probs -> H * W, probs
level_scores = level_probs[:, 1]
level_deltas = tf.reshape(rpn_deltas[idx], [-1, 4])
level_anchors = mlvl_anchors[idx]
pre_nms_limit = tf.math.minimum(num_pre_nms,
tf.shape(level_anchors)[0])
ix = tf.nn.top_k(level_scores, pre_nms_limit, sorted=False).indices
level_scores = tf.gather(level_scores, ix)
level_deltas = tf.gather(level_deltas, ix)
level_anchors = tf.gather(level_anchors, ix)
mlvl_scores.append(level_scores)
mlvl_deltas.append(level_deltas)
mlvl_valid_anchors.append(level_anchors)
level_ids.append(tf.fill([
tf.shape(level_scores)[0],
], idx))
scores = tf.concat(mlvl_scores, axis=0)
anchors = tf.concat(mlvl_valid_anchors, axis=0)
deltas = tf.concat(mlvl_deltas, axis=0)
# get refined anchors
proposals = transforms.delta2bbox(anchors, deltas, self.target_means,
self.target_stds)
# Clip to valid area
window = tf.stack([0., 0., img_shape[0], img_shape[1]])
proposals = transforms.bbox_clip(proposals, window)
ids = tf.concat(level_ids, axis=0)
# NMS is appied per level independent of others
keep = self.batched_nms(proposals, scores, ids, proposal_count,
self.nms_threshold)
proposals = tf.gather(proposals, keep)
return tf.stop_gradient(proposals)
def call(self, inputs, training=True):
'''
Args
---
proposals_list: List of Tensors of shape [num_proposals, (ymin, xmin, ymax, xmax)]
num_proposals=levels * proposals per level. levels refer to FPN levels.
Length of list is the batch size
gt_boxes: Tensor of shape [batch_size, 4]
gt_class_ids: Tensor of shape [batch_size]
img_metas: Tensor of shape [11]
rcnn_feature_maps: List of outputs from the FPN
'''
if training:
proposals_list, rcnn_feature_maps, gt_boxes, \
gt_class_ids, img_metas = inputs
else:
proposals_list, rcnn_feature_maps, img_metas = inputs
batch_size = img_metas.shape[0]
loss_dict = {}
for i in range(self.num_stages):
if i == 0:
rois_list = proposals_list
if training:
rois_list, rcnn_target_matches, rcnn_target_deltas, inside_weights, \
outside_weights = self.bbox_targets[i].build_targets( \
rois_list, gt_boxes, gt_class_ids, img_metas)
pooled_regions_list = self.bbox_roi_extractor(
(rois_list, rcnn_feature_maps, img_metas), training=training)
rcnn_class_logits, rcnn_probs, rcnn_deltas = self.bbox_heads[i](pooled_regions_list, training=training)
if training:
loss_dict['rcnn_class_loss_stage_{}'.format(i)] = losses.rcnn_class_loss(rcnn_class_logits,
rcnn_target_matches) * self.stage_loss_weights[i]
loss_dict['rcnn_box_loss_stage_{}'.format(i)] = losses.rcnn_bbox_loss(rcnn_deltas,
rcnn_target_deltas,
inside_weights,
outside_weights) * self.stage_loss_weights[i]
roi_shapes = [tf.shape(i)[0] for i in rois_list]
refinements = tf.split(rcnn_deltas, roi_shapes)
new_rois = []
if i<(self.num_stages-1):
for j in range(batch_size):
new_rois.append(tf.stop_gradient(transforms.delta2bbox(rois_list[j], refinements[j],
target_means=self.bbox_heads[i].target_means, \
target_stds=self.bbox_heads[i].target_stds)))
rois_list = new_rois
if training:
return loss_dict
else:
detections_list = self.bbox_heads[-1].get_bboxes(rcnn_probs,
rcnn_deltas,
rois_list,
img_metas)
detections_dict = {
'bboxes': detections_list[0][0],
'labels': detections_list[0][1],
'scores': detections_list[0][2]
}
return detections_dict
def _get_bboxes_single(self, rcnn_probs, rcnn_deltas, rois, img_shape):
'''
Args
---
rcnn_probs: [num_rois, num_classes]
rcnn_deltas: [num_rois, num_classes, (dy, dx, log(dh), log(dw))]
rois: [num_rois, (y1, x1, y2, x2)]
img_shape: np.ndarray. [2]. (img_height, img_width)
'''
H = img_shape[0]
W = img_shape[1]
res_scores = tf.TensorArray(tf.float32, size=0, dynamic_size=True, infer_shape=True)
res_bboxes = tf.TensorArray(tf.float32, size=0, dynamic_size=True, infer_shape=True)
res_cls = tf.TensorArray(tf.int32, size=0, dynamic_size=True, infer_shape=True)
for cls_id in range(1, self.num_classes):
inds = tf.where(rcnn_probs[:, cls_id] > self.min_confidence)[:, 0]
cls_score = tf.gather(rcnn_probs[:, cls_id], inds)
rcnn_deltas = tf.reshape(rcnn_deltas, [-1, self.num_classes, 4])
final_bboxes = transforms.delta2bbox(tf.gather(rois, inds),
tf.gather(rcnn_deltas[:, cls_id, :], inds),
self.target_means, self.target_stds)
window = tf.stack([tf.constant(0., H.dtype), tf.constant(0., W.dtype), H, W])
final_bboxes = transforms.bbox_clip(final_bboxes, window)
cls_score = tf.cast(cls_score, final_bboxes.dtype)
#keep = tf.image.non_max_suppression(final_bboxes, cls_score,
# self.max_instances,
# iou_threshold=self.nms_threshold)
keep, selected_cls_scores, _ = tf.raw_ops.NonMaxSuppressionV5 (
boxes=final_bboxes, scores=cls_score,
max_output_size=self.max_instances,
iou_threshold=self.nms_threshold,
score_threshold=0.0,
soft_nms_sigma=self.soft_nms_sigma)
pad_size = self.max_instances - tf.size(keep)
padded_scores = tf.pad(selected_cls_scores, paddings=[[0, pad_size]], constant_values=0.0)
res_scores = res_scores.write(cls_id-1, padded_scores)#.mark_used()
padded_bboxes = tf.pad(tf.gather(final_bboxes, keep), paddings=[[0, pad_size], [0, 0]], constant_values=0.0)
res_bboxes = res_bboxes.write(cls_id-1, padded_bboxes)#.mark_used()
padded_cls = tf.pad(tf.ones_like(keep, dtype=tf.int32) * cls_id, paddings=[[0, pad_size]], constant_values=-1)
res_cls = res_cls.write(cls_id-1, padded_cls)#.mark_used()
res_scores = res_scores.stack()
res_bboxes = res_bboxes.stack()
res_cls = res_cls.stack()
scores_after_nms = tf.reshape(res_scores, [-1])
bboxes_after_nms = tf.reshape(res_bboxes, [-1, 4])
cls_after_nms = tf.reshape(res_cls, [-1])
_, final_idx = tf.nn.top_k(scores_after_nms,
k=tf.minimum(self.max_instances, tf.size(scores_after_nms)),
sorted=False)
return (tf.gather(bboxes_after_nms, final_idx),
tf.gather(cls_after_nms, final_idx),
tf.gather(scores_after_nms, final_idx))
def _get_proposals_single(self,
scores_list,
deltas_list,
anchors_list,
img_shape,
with_probs,
training=True):
"""
Transform outputs for a single batch item into labeled boxes.
"""
assert len(scores_list) == len(deltas_list) == len(anchors_list)
level_ids = []
mlvl_deltas = []
mlvl_scores = []
mlvl_anchors = []
mlvl_proposals = []
num_levels = len(scores_list)
for idx in range(num_levels):
probs = tf.keras.layers.Activation(
tf.nn.sigmoid, dtype=tf.float32)(scores_list[idx])
deltas = deltas_list[idx]
anchors = anchors_list[idx]
probs = tf.reshape(probs, [-1, self.num_classes])
deltas = tf.reshape(deltas, [-1, 4])
pre_nms_limit = tf.math.minimum(self.num_pre_nms,
tf.shape(anchors)[0])
max_probs = tf.reduce_max(probs, axis=1)
ix = tf.nn.top_k(
max_probs,
k=pre_nms_limit).indices # top k for each level (as per paper)
level_anchors = tf.gather(anchors, ix)
level_deltas = tf.gather(deltas, ix)
level_scores = tf.gather(probs, ix) # these contain max_probs
mlvl_deltas.append(level_deltas)
mlvl_scores.append(level_scores)
mlvl_anchors.append(level_anchors)
scores = tf.concat(mlvl_scores, axis=0)
anchors = tf.concat(mlvl_anchors, axis=0)
deltas = tf.concat(mlvl_deltas, axis=0)
proposals = transforms.delta2bbox(anchors, deltas, self.target_means,
self.target_stds)
# Clip to valid area
window = tf.stack([0., 0., img_shape[0], img_shape[1]])
proposals = transforms.bbox_clip(proposals, window)
return self.batched_nms(proposals, scores, self.max_instances,
self.nms_threshold)
def _get_proposals_single(self,
rpn_probs,
rpn_deltas,
anchors,
valid_flags,
img_shape,
with_probs,
training=True):
'''
Calculate proposals.
Args
---
rpn_probs: [num_anchors]
rpn_deltas: [num_anchors, (dy, dx, log(dh), log(dw))]
anchors: [num_anchors, (y1, x1, y2, x2)] anchors defined in
pixel coordinates.
valid_flags: [num_anchors]
img_shape: np.ndarray. [2]. (img_height, img_width)
with_probs: bool.
Returns
---
proposals: [num_proposals, (y1, x1, y2, x2)] in normalized
coordinates.
'''
H = img_shape[0]
W = img_shape[1]
# filter invalid anchors, int => bool
valid_flags = tf.cast(valid_flags, tf.bool)
rpn_probs = tf.boolean_mask(rpn_probs, valid_flags)
rpn_deltas = tf.boolean_mask(rpn_deltas, valid_flags)
anchors = tf.boolean_mask(anchors, valid_flags)
# Improve performance
if training:
num_pre_nms = self.num_pre_nms_train
else:
num_pre_nms = self.num_pre_nms_test
pre_nms_limit = tf.math.minimum(num_pre_nms, tf.shape(anchors)[0])
ix = tf.nn.top_k(rpn_probs, pre_nms_limit, sorted=False).indices
rpn_probs = tf.gather(rpn_probs, ix)
rpn_deltas = tf.gather(rpn_deltas, ix)
anchors = tf.gather(anchors, ix)
# Get refined anchors
proposals = transforms.delta2bbox(anchors, rpn_deltas,
self.target_means, self.target_stds)
# Clip to valid area
window = tf.stack([0., 0., H, W])
proposals = transforms.bbox_clip(proposals, window)
if training:
proposal_count = self.num_post_nms_train
else:
proposal_count = self.num_post_nms_test
rpn_probs = tf.cast(rpn_probs, proposals.dtype)
# indices = tf.image.non_max_suppression(proposals, rpn_probs,
# max_output_size=proposal_count,
# iou_threshold=self.nms_threshold)
indices = tf.raw_ops.NonMaxSuppressionV2(
boxes=proposals,
scores=rpn_probs,
max_output_size=proposal_count,
iou_threshold=self.nms_threshold)
proposals = tf.stop_gradient(tf.gather(proposals, indices))
if with_probs:
proposal_probs = tf.expand_dims(tf.gather(rpn_probs, indices),
axis=1)
proposals = tf.concat([proposals, proposal_probs], axis=1)
return proposals
