def reshape(self, bottom, top):
set_blob_data(
top[0],
np.zeros((1, self.ver_num, self.hor_num, 2), dtype=np.float32))
set_blob_data(
top[1],
np.zeros((1, self.ver_num, self.hor_num, 1), dtype=np.float32))
def reshape(self, bottom, top):
set_blob_data(top[0], np.zeros((1, 9), dtype=np.float32))
set_blob_data(top[1], np.zeros((MAX_LEN, 1), dtype=np.float32))
set_blob_data(top[2], np.zeros((MAX_LEN, 1), dtype=np.float32))
set_blob_data(top[3], np.zeros((MAX_LEN, 1), dtype=np.float32))
if len(top) > 4:
set_blob_data(top[4], np.zeros((1, 5), dtype=np.float32))
def forward(self, bottom, top):
ps_embedding_blob = np.zeros_like(bottom[0].data, dtype=np.float32)
for t in range(bottom[0].data.shape[0]):
for n in range(bottom[0].data.shape[1]):
id_x = np.argmax(bottom[0].data[t, n])
ps_embedding_blob[t, n, id_x] = 1
set_blob_data(top[0], ps_embedding_blob)
def forward(self, bottom, top):
if len(bottom) > 3:
self.fcn_th = float(bottom[3].data[0, 0])
pre_score = bottom[0].data[0, 1, :, :]
pre_bbox = bottom[1].data[0, :, :, :]
pre_orient = bottom[2].data[0, 0, :, :]
pre_bbox *= self.rf
pos_y, pos_x = np.where(pre_score > self.fcn_th)
bboxes = np.zeros((len(pos_y), 7), dtype='float32')
orients = np.zeros((len(pos_y)), dtype='float32')
for i in range(len(pos_y)):
y = pos_y[i]
x = pos_x[i]
t, b, l, r = pre_bbox[:, y, x]
bboxes[i] = np.asarray(
(x - l, y - t, x + r, y + b, pre_score[y, x], x, y))
orients[i] = pre_orient[y, x]
orient_bboxes = np.zeros((len(bboxes), 11), dtype='float32')
for i, box in enumerate(bboxes):
box[0] *= 4
box[2] *= 4
box[1] *= 4
box[3] *= 4
box[5] *= 4
box[6] *= 4
temp_box = rotate_rect(box[0], box[1], box[2], box[3], orients[i],
box[5], box[6])
orient_bboxes[i] = np.array(
(temp_box[0][0], temp_box[0][1], temp_box[1][0],
temp_box[1][1], temp_box[2][0], temp_box[2][1],
temp_box[3][0], temp_box[3][1], box[4], box[-2], box[-1]))
keep_indices, temp_boxes = non_max_suppression(orient_bboxes,
self.nms_th)
#assert len(keep_indices)==len(temp_boxes)
#orient_bboxes = temp_boxes[keep_indices]
orient_bboxes = orient_bboxes[keep_indices]
if len(orient_bboxes) > 0:
set_blob_data(
top[0],
np.array(np.hstack((np.zeros(
(orient_bboxes.shape[0], 1)), orient_bboxes[:, :9])),
dtype=np.float32))
if len(top) > 1:
cont = np.ones((25, orient_bboxes.shape[0]), dtype=np.float32)
cont[0, :] = 0
set_blob_data(top[1], cont)
else:
set_blob_data(top[0], np.zeros((1, 10), dtype=np.float32))
if len(top) > 1:
set_blob_data(top[1], np.zeros((25, 1), dtype=np.float32))
def forward(self, bottom, top):
roi_num = bottom[0].data.shape[0]
rois = bottom[0].data
sample_points = np.zeros((roi_num, self.ver_num, self.hor_num, 2))
sample_id = np.zeros((roi_num, self.ver_num, self.hor_num, 1))
for k in range(roi_num):
bb = rois[k, 1:]
edge1 = ((bb[0] - bb[2])**2 + (bb[1] - bb[3])**2)
edge2 = ((bb[0] - bb[6])**2 + (bb[1] - bb[7])**2)
#angle1 = atan(abs(bb[1] - bb[3]) / abs(bb[0] - bb[2]))
#angle2 = atan(abs(bb[1] - bb[7]) / abs(bb[0] - bb[6]))
point1 = np.zeros(2)
point2 = np.zeros(2)
box = bb
if edge1 < edge2:
for i in range(self.ver_num):
point1[0] = box[0] + (box[2] - box[0]) / float(
self.ver_num + 1) * float(i + 1)
point1[1] = box[1] + (box[3] - box[1]) / float(
self.ver_num + 1) * float(i + 1)
point2[0] = box[6] + (box[4] - box[6]) / float(
self.ver_num + 1) * float(i + 1)
point2[1] = box[7] + (box[5] - box[7]) / float(
self.ver_num + 1) * float(i + 1)
if point1[0] < point2[0]:
sample_points[k, i, :,
0] = np.linspace(point1[0], point2[0],
self.hor_num)
sample_points[k, i, :,
1] = np.linspace(point1[1], point2[1],
self.hor_num)
else:
sample_points[k, i, :,
0] = np.linspace(point2[0], point1[0],
self.hor_num)
sample_points[k, i, :,
1] = np.linspace(point2[1], point1[1],
self.hor_num)
else:
for i in range(self.ver_num):
point1[0] = box[0] + (box[6] - box[0]) / float(
self.ver_num + 1) * float(i + 1)
point1[1] = box[1] + (box[7] - box[1]) / float(
self.ver_num + 1) * float(i + 1)
point2[0] = box[2] + (box[4] - box[2]) / float(
self.ver_num + 1) * float(i + 1)
point2[1] = box[3] + (box[5] - box[3]) / float(
self.ver_num + 1) * float(i + 1)
if point1[0] < point2[0]:
sample_points[k, i, :,
0] = np.linspace(point1[0], point2[0],
self.hor_num)
sample_points[k, i, :,
1] = np.linspace(point1[1], point2[1],
self.hor_num)
else:
sample_points[k, i, :,
0] = np.linspace(point2[0], point1[0],
self.hor_num)
sample_points[k, i, :,
1] = np.linspace(point2[1], point1[1],
self.hor_num)
sample_id[k] = rois[k, 0]
set_blob_data(top[0], sample_points)
set_blob_data(top[1], sample_id)
def forward(self, bottom, top):
batch_size = bottom[0].data.shape[0]
# gt_boxes = []
cont = []
input_label = []
output_label = []
gt_boxes = np.zeros((0, 9))
for n in range(batch_size):
gt_label = bottom[0].data[n, 0]
tmp = np.sum(gt_label, axis=1)
gt_num = len(np.where(tmp != 0)[0])
if gt_num == 0:
continue
roi_n = gt_label[:gt_num, :8] * 4
roi_n = np.hstack((np.ones((gt_num, 1)) * n, roi_n))
gt_boxes = np.vstack((gt_boxes, roi_n))
for k in range(gt_num):
label_len = int(gt_label[k, 9])
if label_len > MAX_LEN - 1:
cont.append([0] * MAX_LEN)
input_label.append([0] * MAX_LEN)
output_label.append([-1] * MAX_LEN)
continue
pad = MAX_LEN - label_len - 1
cont_tmp = [0] + [1] * label_len + [0] * pad
cont.append(cont_tmp)
input_tmp = [0] + list(
gt_label[k, 10:10 + label_len]) + [-1] * pad
input_label.append(input_tmp)
output_tmp = list(
gt_label[k, 10:10 + label_len]) + [0] + [-1] * pad
output_label.append(output_tmp)
if len(gt_boxes) == 0:
set_blob_data(top[0], np.zeros((1, 9), dtype=np.float32))
set_blob_data(top[1], np.zeros((MAX_LEN, 1), dtype=np.float32))
set_blob_data(top[2], np.zeros((MAX_LEN, 1), dtype=np.float32))
set_blob_data(top[3], -np.ones((MAX_LEN, 1), dtype=np.float32))
if len(top) > 4:
set_blob_data(top[4], np.zeros((1, 5), dtype=np.float32))
return
gt_boxes = np.array(gt_boxes).reshape(-1, 9)
cont = np.array(cont,
dtype=np.float32).reshape(-1, MAX_LEN).transpose(1, 0)
input_label = np.array(input_label, dtype=np.float32).reshape(
-1, MAX_LEN).transpose(1, 0)
output_label = np.array(output_label, dtype=np.float32).reshape(
-1, MAX_LEN).transpose(1, 0)
gt_len = gt_boxes.shape[0]
sel_num = min(gt_len, self.sel_num)
sel_ids = np.int32(
np.random.choice(np.arange(gt_len), sel_num, replace=False))
gt_boxes = gt_boxes[sel_ids]
cont = cont[:, sel_ids]
input_label = input_label[:, sel_ids]
output_label = output_label[:, sel_ids]
rects = np.zeros((gt_boxes.shape[0], 5), dtype=np.float32)
for nn in range(gt_boxes.shape[0]):
rects[nn, 0] = gt_boxes[nn, 0]
rects[nn, 1] = np.min(gt_boxes[nn, [1, 3, 5, 7]])
rects[nn, 2] = np.min(gt_boxes[nn, [2, 4, 6, 8]])
rects[nn, 3] = np.max(gt_boxes[nn, [1, 3, 5, 7]])
rects[nn, 4] = np.max(gt_boxes[nn, [2, 4, 6, 8]])
if self.phase == 0:
cont = np.ones((25, int(cont.shape[1])), dtype=np.float32)
cont[0, :] = 0
set_blob_data(top[0], gt_boxes)
set_blob_data(top[1], cont)
set_blob_data(top[2], input_label)
set_blob_data(top[3], output_label)
if len(top) > 4:
set_blob_data(top[4], rects)
def reshape(self, bottom, top):
set_blob_data(top[0], np.zeros((1, 10), dtype=np.float32))
if len(top) > 1:
set_blob_data(top[1], np.zeros((25, 1), dtype=np.float32))
def reshape(self, bottom, top):
set_blob_data(top[0], np.zeros_like(bottom[0].data, dtype=np.float32))
def forward(self, bottom, top):
roi_num = bottom[0].data.shape[0]
assert (roi_num > 0)
rois = bottom[0].data
sample_points = np.zeros((roi_num, self.ver_num, self.hor_num, 2))
sample_id = np.zeros((roi_num, self.ver_num, self.hor_num, 1))
for k in range(roi_num):
bb = rois[k, 1:]
edge1 = ((bb[0] - bb[2])**2 + (bb[1] - bb[3])**2)
edge2 = ((bb[0] - bb[6])**2 + (bb[1] - bb[7])**2)
angle1 = abs(atan((bb[1] - bb[3]) / (bb[0] - bb[2] + 0.001)))
angle2 = abs(atan((bb[1] - bb[7]) / (bb[0] - bb[6] + 0.001)))
point1 = np.zeros(2)
point2 = np.zeros(2)
box = bb
#print angle1,angle2
if angle1 > angle2:
for i in range(self.ver_num):
point1[0] = box[0] + (box[2] - box[0]) / float(
self.ver_num + 1) * float(i + 1)
point1[1] = box[1] + (box[3] - box[1]) / float(
self.ver_num + 1) * float(i + 1)
point2[0] = box[6] + (box[4] - box[6]) / float(
self.ver_num + 1) * float(i + 1)
point2[1] = box[7] + (box[5] - box[7]) / float(
self.ver_num + 1) * float(i + 1)
if point1[0] < point2[0]:
sample_points[k, i, :,
0] = np.linspace(point1[0], point2[0],
self.hor_num)
sample_points[k, i, :,
1] = np.linspace(point1[1], point2[1],
self.hor_num)
else:
sample_points[k, i, :,
0] = np.linspace(point2[0], point1[0],
self.hor_num)
sample_points[k, i, :,
1] = np.linspace(point2[1], point1[1],
self.hor_num)
else:
for i in range(self.ver_num):
point1[0] = box[0] + (box[6] - box[0]) / float(
self.ver_num + 1) * float(i + 1)
point1[1] = box[1] + (box[7] - box[1]) / float(
self.ver_num + 1) * float(i + 1)
point2[0] = box[2] + (box[4] - box[2]) / float(
self.ver_num + 1) * float(i + 1)
point2[1] = box[3] + (box[5] - box[3]) / float(
self.ver_num + 1) * float(i + 1)
if point1[0] < point2[0]:
sample_points[k, i, :,
0] = np.linspace(point1[0], point2[0],
self.hor_num)
sample_points[k, i, :,
1] = np.linspace(point1[1], point2[1],
self.hor_num)
else:
sample_points[k, i, :,
0] = np.linspace(point2[0], point1[0],
self.hor_num)
sample_points[k, i, :,
1] = np.linspace(point2[1], point1[1],
self.hor_num)
sample_id[k] = rois[k, 0]
set_blob_data(top[0], sample_points)
set_blob_data(top[1], sample_id)
if len(top) > 2:
set_blob_data(top[2], np.mean(sample_points, axis=1))
set_blob_data(top[3], np.mean(sample_id, axis=1))