def _rezoom(hyp, pred_boxes, early_feat, early_feat_channels,
w_offsets, h_offsets):
#Rezoom into a feature map at multiple interpolation points
#in a grid.
#If the predicted object center is at X, len(w_offsets) == 3,
#and len(h_offsets) == 5,
#the rezoom grid will look as follows:
#[o o o]
#[o o o]
#[o X o]
#[o o o]
#[o o o]
#Where each letter indexes into the feature map with bilinear interpolation
#grid size
grid_size = hyp['grid_width'] * hyp['grid_height']
outer_size = grid_size * hyp['batch_size']
indices = []
for w_offset in w_offsets:
for h_offset in h_offsets:
indices.append(train_utils.bilinear_select(hyp,
pred_boxes,
early_feat,
early_feat_channels,
w_offset, h_offset))
#Rezoom into a feature map at multiple interpolation points
interp_indices = tf.concat(axis=0, values=indices)
rezoom_features = train_utils.interp(early_feat,
interp_indices,
early_feat_channels)
def rezoom(H, pred_boxes, early_feat, early_feat_channels, w_offsets,
h_offsets):
'''
Rezoom into a feature map at multiple interpolation points in a grid.
If the predicted object center is at X, len(w_offsets) == 3, and len(h_offsets) == 5,
the rezoom grid will look as follows:
[o o o]
[o o o]
[o X o]
[o o o]
[o o o]
Where each letter indexes into the feature map with bilinear interpolation
'''
grid_size = H['grid_width'] * H['grid_height']
outer_size = grid_size * H['batch_size']
indices = []
for w_offset in w_offsets:
for h_offset in h_offsets:
indices.append(
train_utils.bilinear_select(H, pred_boxes, early_feat,
early_feat_channels, w_offset,
h_offset))
interp_indices = tf.concat(axis=0, values=indices)
rezoom_features = train_utils.interp(early_feat, interp_indices,
early_feat_channels)
rezoom_features_r = tf.reshape(rezoom_features, [
len(w_offsets) * len(h_offsets), outer_size, H['rnn_len'],
early_feat_channels
])
rezoom_features_t = tf.transpose(rezoom_features_r, [1, 2, 0, 3])
return tf.reshape(rezoom_features_t, [
outer_size, H['rnn_len'],
len(w_offsets) * len(h_offsets) * early_feat_channels
])
def rezoom(H, pred_boxes, early_feat, early_feat_channels, w_offsets, h_offsets):
'''
Rezoom into a feature map at multiple interpolation points in a grid.
If the predicted object center is at X, len(w_offsets) == 3, and len(h_offsets) == 5,
the rezoom grid will look as follows:
[o o o]
[o o o]
[o X o]
[o o o]
[o o o]
Where each letter indexes into the feature map with bilinear interpolation
'''
grid_size = H['grid_width'] * H['grid_height']
outer_size = grid_size * H['batch_size']
indices = []
for w_offset in w_offsets:
for h_offset in h_offsets:
indices.append(train_utils.bilinear_select(H,
pred_boxes,
early_feat,
early_feat_channels,
w_offset, h_offset))
interp_indices = tf.concat(0, indices)
rezoom_features = train_utils.interp(early_feat,
interp_indices,
early_feat_channels)
rezoom_features_r = tf.reshape(rezoom_features,
[len(w_offsets) * len(h_offsets),
outer_size,
H['rnn_len'],
early_feat_channels])
rezoom_features_t = tf.transpose(rezoom_features_r, [1, 2, 0, 3])
return tf.reshape(rezoom_features_t,
[outer_size,
H['rnn_len'],
len(w_offsets) * len(h_offsets) * early_feat_channels])
def rezoom(H, pred_boxes, early_feat, early_feat_channels, w_offsets,
h_offsets):
grid_size = H['grid_width'] * H['grid_height']
outer_size = grid_size * H['batch_size']
indices = []
for w_offset in w_offsets:
for h_offset in h_offsets:
indices.append(
train_utils.bilinear_select(H, pred_boxes, early_feat,
early_feat_channels, w_offset,
h_offset))
interp_indices = tf.concat(axis=0, values=indices)
rezoom_features = train_utils.interp(early_feat, interp_indices,
early_feat_channels)
rezoom_features_r = tf.reshape(rezoom_features, [
len(w_offsets) * len(h_offsets), outer_size, H['rnn_len'],
early_feat_channels
])
rezoom_features_t = tf.transpose(rezoom_features_r, [1, 2, 0, 3])
return tf.reshape(rezoom_features_t, [
outer_size, H['rnn_len'],
len(w_offsets) * len(h_offsets) * early_feat_channels
])