def riounms(pathsrc, classes, thresh):
file = open(pathsrc, 'r').readlines()
os.remove(pathsrc)
bboxes = []
for bb in file:
bb = bb.split()
score, cl, x1, y1, x2, y2, x3, y3, x4, y4 = float(bb[0]), float(
classes.index(str(bb[1]))), float(bb[2]), float(bb[3]), float(
bb[4]), float(bb[5]), float(bb[6]), float(bb[7]), float(
bb[8]), float(bb[9])
bboxes.append([score, cl, x1, y1, x2, y2, x3, y3, x4, y4])
bboxes = np.array(bboxes).astype('float32')
bboxfiltered = []
#bboxes = bboxes[bboxes[:,0]>0.05]
for c in range(15):
dets = bboxes[bboxes[:, 1] == c]
if dets.shape[0] == 0:
continue
scores = dets[:, 0]
polys = []
areas = []
for i in range(len(dets)):
tm_polygon = polyiou.VectorDouble([
float(dets[i][2]),
float(dets[i][3]),
float(dets[i][4]),
float(dets[i][5]),
float(dets[i][6]),
float(dets[i][7]),
float(dets[i][8]),
float(dets[i][9])
])
polys.append(tm_polygon)
order = scores.argsort()[::-1]
keep = []
while order.size > 0:
ovr = []
i = order[0]
keep.append(i)
for j in range(order.size - 1):
iou = polyiou.iou_poly(polys[i], polys[order[j + 1]])
ovr.append(iou)
ovr = np.array(ovr)
inds = np.where(ovr <= thresh)[0]
order = order[inds + 1]
bboxfiltered.append(dets[keep])
if len(bboxfiltered) == 0:
return
bboxfiltered = np.concatenate(bboxfiltered, axis=0)
with open(pathsrc, 'w') as fin:
for box in bboxfiltered:
score, cl, x1, y1, x2, y2, x3, y3, x4, y4 = box[0], classes[int(
box[1]
)], box[2], box[3], box[4], box[5], box[6], box[7], box[8], box[9]
fin.write(
str(score)[0:5] + ' ' + cl + ' ' + str(int(x1)) + ' ' +
str(int(y1)) + ' ' + str(int(x2)) + ' ' + str(int(y2)) + ' ' +
str(int(x3)) + ' ' + str(int(y3)) + ' ' + str(int(x4)) + ' ' +
str(int(y4)) + '\n')
def calcoverlaps(BBGT_keep, bb):
overlaps = []
for index, GT in enumerate(BBGT_keep):
overlap = polyiou.iou_poly(
polyiou.VectorDouble(BBGT_keep[index]),
polyiou.VectorDouble(bb))
overlaps.append(overlap)
return overlaps
def riounms(bboxes_per_img, classes, thresh=0.2):
if len(bboxes_per_img) == 0:
return []
bboxes = []
for bb in bboxes_per_img:
x1,y1,x2,y2,x3,y3,x4,y4 = bb['bbox']
score,cl,x1,y1,x2,y2,x3,y3,x4,y4 = float(bb['score']),float(classes.index(bb['class'])),float(x1),float(y1),float(x2),float(y2),float(x3),float(y3),float(x4),float(y4)
bboxes.append([score,cl,x1,y1,x2,y2,x3,y3,x4,y4])
bboxes = np.array(bboxes).astype('float32')
bboxfiltered = []
for c in range(15):
dets = bboxes[bboxes[:,1]==c]
if dets.shape[0] == 0:
continue
scores = dets[:, 0]
polys = []
areas = []
for i in range(len(dets)):
tm_polygon = polyiou.VectorDouble([float(dets[i][2]), float(dets[i][3]), float(dets[i][4]), float(dets[i][5]), float(dets[i][6]), float(dets[i][7]), float(dets[i][8]), float(dets[i][9])])
polys.append(tm_polygon)
order = scores.argsort()[::-1]
keep = []
while order.size > 0:
ovr = []
i = order[0]
keep.append(i)
for j in range(order.size - 1):
iou = polyiou.iou_poly(polys[i], polys[order[j + 1]])
ovr.append(iou)
ovr = np.array(ovr)
inds = np.where(ovr <= thresh)[0]
order = order[inds + 1]
bboxfiltered.append(dets[keep])
if len(bboxfiltered)==0:
return []
bboxfiltered = np.concatenate(bboxfiltered,axis=0)
bboxes_out = []
for box in bboxfiltered:
bb={}
bb['bbox']=[box[2],box[3],box[4],box[5],box[6],box[7],box[8],box[9]]
bb['score']=box[0]
bb['class']=classes[int(box[1])]
bboxes_out.append(bb)
return bboxes_out
def forward(self, classifications, regressions, anchors, annotations,
**kwargs):
alpha = 0.25
gamma = 2.0
# epsilon = 0.15 # used for label smoothing
batch_size = classifications.shape[0]
classification_losses = []
regression_losses = []
anchor = anchors[0, :, :] # shape:[xmin, ymin, xmax, ymax, theta]
dtype = anchors.dtype
anchor_widths = anchor[:, 2] - anchor[:, 0]
anchor_heights = anchor[:, 3] - anchor[:, 1]
anchor_ctr_x = anchor[:, 0] + 0.5 * anchor_widths
anchor_ctr_y = anchor[:, 1] + 0.5 * anchor_heights
anchor_theta = anchor[:, 4]
for j in range(batch_size):
classification = classifications[j, :, :]
regression = regressions[j, :, :]
bbox_annotation = annotations[
j] # bbox_annotation -->(x_c, y_c, width, height, theta, cls)
bbox_annotation = bbox_annotation[bbox_annotation[:, 5] != -1]
# log(prob) if prob is too small, the log(prob) will be Nan
classification = torch.clamp(classification, 1e-4, 1.0 - 1e-4)
if bbox_annotation.shape[0] == 0:
if torch.cuda.is_available():
alpha_factor = torch.ones_like(classification) * alpha
alpha_factor = alpha_factor.cuda()
alpha_factor = 1. - alpha_factor
focal_weight = classification
focal_weight = alpha_factor * torch.pow(
focal_weight, gamma)
bce = -(torch.log(1.0 - classification))
cls_loss = focal_weight * bce
regression_losses.append(torch.tensor(0).to(dtype).cuda())
classification_losses.append(cls_loss.sum())
else:
alpha_factor = torch.ones_like(classification) * alpha
alpha_factor = 1. - alpha_factor
focal_weight = classification
focal_weight = alpha_factor * torch.pow(
focal_weight, gamma)
bce = -(torch.log(1.0 - classification))
cls_loss = focal_weight * bce
regression_losses.append(torch.tensor(0).to(dtype))
classification_losses.append(cls_loss.sum())
continue
# Get positive and negative samples
# Step1. Get horizontal overlaps between anchors and gt box
targets = torch.ones_like(classification) * -1
if torch.cuda.is_available():
targets = targets.cuda()
# 1) 首先获取gt(bbox_annotation)对应的poly坐标
vertex = Rectangle_area(
bbox_annotation[:, :5]) # vertex --> ndarray
# Debug
# visualize_Rectangle_area(image_path, vertex[:5])
# vertex [[[x1, y1], [x2, y2], [x3, y3], [x4, y4]], ... [[], [], [], []]]
# 2) 将poly坐标转换成最小的外接水平包围框
horizontal_vertex = poly2Horizontal_rect(
vertex) # horizontal_vertex --> ndarray
# Debug
# visualize_poly2Horizontal_rect(image_path, horizontal_vertex[:5])
# horizontal_vertex [[xmin, ymin, xmax, ymax], ... ,[]]
# 3) 计算最小的外接水平包围框与gt(bbox_annotation)的horizontal overlaps
if torch.cuda.is_available():
horizontal_vertex = torch.tensor(horizontal_vertex).cuda()
HIoU = calc_iou(
anchor[:, :], horizontal_vertex
) # HIoU shape (torch.Size([len(anchor), len(anno)])
hor_IoU_max, hor_IoU_argmax = torch.max(HIoU, dim=1)
# Step2. 根据设定的水平框的IoU阈值(hor_threshold = 0.2)，获得可能使正样本的Anchor index
hor_positive_indices = torch.ge(hor_IoU_max, 0.6)
# print(f"第一次水平框筛选后：水平正类的个数为{hor_positive_indices.sum()}")
parent_num_list = np.arange(len(anchor))
parent_positive_index = list(
parent_num_list[hor_positive_indices.cpu().numpy()])
positive_anchor_list = anchor[hor_positive_indices, :]
# Debug
# visualize_positive_anchor(image_path, anchor, parent_positive_index)
# print(temp_parent_positive_index)
# parent_positive_index = find_index(hor_positive_indices) # 原始代码，速度很慢，用parent_num_list[]进行代替
# print(f"line 188: 经过水平框的初步筛选，水平正类有:{hor_positive_indices.sum()}")
# 4、将IOU大于设定的hor_threshold的水平包围框对应的旋转框与gt计算skew IOU，然后再计算正类
# 将所有的Anchor先将gt进行初步分配
assigned_annotations = bbox_annotation[hor_IoU_argmax, :]
# 将可能为正类样本的Anchor分配与其匹配的gt信息(x1, y1, width, height, theta, cls)
# 此时hor_positive_assigned_annotations存放与anchor相匹配的gt信息
hor_positive_assigned_annotations = assigned_annotations[
hor_positive_indices, :]
# print(f"line 199: 当前每个Anchor对应的annotation:{hor_positive_assigned_annotations.size()}")
# 获取可能正类样本的Anchor自身坐标及宽高信息用于计算Anchor与gt的skew IoU
# print(f"line 206: -------->{hor_positive_indices}")
# 目前hor_positive_indices是tensor，但是anchor_widths_pi等四个变量均是ndarray
# hor_positive_indices = hor_positive_indices.cpu().numpy()
anchor_widths_pi = anchor_widths[
hor_positive_indices] # shape(len(hor_positive_indices), )
anchor_heights_pi = anchor_heights[hor_positive_indices]
anchor_ctr_x_pi = anchor_ctr_x[hor_positive_indices]
anchor_ctr_y_pi = anchor_ctr_y[hor_positive_indices]
xlt, ylt = anchor_ctr_x_pi - anchor_widths_pi / 2, anchor_ctr_y_pi - anchor_heights_pi / 2
xrt, yrt = anchor_ctr_x_pi + anchor_widths_pi / 2, anchor_ctr_y_pi - anchor_heights_pi / 2
xrb, yrb = anchor_ctr_x_pi + anchor_widths_pi / 2, anchor_ctr_y_pi + anchor_heights_pi / 2
xlb, ylb = anchor_ctr_x_pi - anchor_widths_pi / 2, anchor_ctr_y_pi + anchor_heights_pi / 2
anchor_vertex = torch.stack(
[xlt, ylt, xrt, yrt, xrb, yrb, xlb, ylb]).t()
# 可视化anchor_vertex表示的是否正确
# check_anchor(image_path, anchor_vertex)
# 计算可能为正样本的Anchor与gt的 skew IoU Note:anchor_vertex中anchor[i] 与
# hor_positive_assigned_annotations[i] 是一一对应的关系
skew_IoU_lists = []
for index in range(len(anchor_vertex)):
single_anchor = anchor_vertex[index]
single_rotation_opencv_format = hor_positive_assigned_annotations[
index][:5]
single_rotation_box = single_Rectangle_area(
single_rotation_opencv_format)
single_points = Rotation2points(single_rotation_box)
# example input for polyiou.iou_poly() function
# temp1 = [433.16364, 186.74037, 421.0, 172.99998, 482.83636, 118.2596, 495.0, 131.99998]
# temp2 = [282.601, 114.6088, 333.39912, 114.60088, 333.394912, 165.912, 282.6088, 165.912]
# 由于上面代码中的single_points, single_anchor对应的类型是np.float64类型，函数中需要float类型
result1 = list(map(lambda x: float(x), single_points))
result2 = list(map(lambda x: float(x), single_anchor))
skew_IoU = polyiou.iou_poly(polyiou.VectorDouble(result1),
polyiou.VectorDouble(result2))
skew_IoU = np.array(skew_IoU).astype(np.float64)
skew_IoU_lists.append(skew_IoU)
skew_IoU_lists = np.array(skew_IoU_lists)
if not torch.is_tensor(skew_IoU_lists):
overlaps = torch.from_numpy(skew_IoU_lists).cuda(0)
# print(f"line 285: -----> {torch.max(overlaps)}")
rotation_threshold = 0.3 # 使用0.2的阈值可以有较高的检测正类个数
rotation_positive_indices = torch.ge(overlaps, rotation_threshold)
# print(f"line 291: -----> {rotation_positive_indices}")
# son_positive_index = find_index(rotation_positive_indices)
son_num_list = np.arange(len(overlaps))
son_positive_index = list(son_num_list[
rotation_positive_indices.cpu().numpy()]) # 对用find_index代码进行优化
# 可视化此时正类Anchor的筛选情况
# print(f"第二次筛选后，Anchor的数量情况{len(son_positive_index)}")
# positive_index_list = np.zeros(len(anchor), dtype=bool)
# for idx in range(len(son_positive_index)):
# son_idx = son_positive_index[idx]
# parent_idx = parent_positive_index[son_idx]
# positive_index_list[parent_idx] = 1
# pp_anchor = anchor[positive_index_list]
#
# visualize_pp_anchor(image_path, pp_anchor)
# target (torch.Tensor): The learning label of the prediction.
# compute the loss for classification
# 对正负样本进行one-hot编码
# 正负样本的划分方式 Zylo117 pos->169 neg->48558
# 初次筛选成功的水平框的anchor为负类，其中通过旋转框筛选的anchor为正类
# targets[hor_positive_indices, :] = 0 # ori code
# targets[torch.lt(IoU_max, 0.4), :] = epsilon # modified for the label smoothing
# temp = torch.lt(hor_IoU_max, 0.4)
# neg_num = temp.sum()
# hor_IoU_max < 0.4 设置为负类样本
targets[torch.lt(hor_IoU_max, 0.4), :] = 0
# skew_IoU > 0.3 设置为正类样本
num_positive_anchors = rotation_positive_indices.sum()
# # 寻找正类样本
# 将positive_index_list 改写为[False, False, True, True, ... False, True]的这种形式
positive_index_list = np.zeros(len(anchor), dtype=bool)
for idx in range(len(son_positive_index)):
son_idx = son_positive_index[idx]
parent_idx = parent_positive_index[son_idx]
positive_index_list[parent_idx] = 1
targets[positive_index_list, :] = 0
targets[positive_index_list,
assigned_annotations[positive_index_list, 5].long()] = 1
# for label smoothing
# targets[positive_indices, :] = epsilon # modified for the label smoothing
# targets[positive_indices, assigned_annotations[positive_indices, 4].long()] = 1.0 - epsilon
alpha_factor = torch.ones_like(targets) * alpha
if torch.cuda.is_available():
alpha_factor = alpha_factor.cuda()
alpha_factor = torch.where(torch.eq(targets, 1.), alpha_factor,
1. - alpha_factor)
focal_weight = torch.where(torch.eq(targets, 1.),
1. - classification, classification)
# for label smoothing
# alpha_factor = torch.where(torch.eq(targets, 1.0 - epsilon), alpha_factor, 1. - alpha_factor)
# focal_weight = torch.where(torch.eq(targets, 1.0 - epsilon), 1. - classification, classification)
focal_weight = alpha_factor * torch.pow(focal_weight, gamma)
bce = -(targets * torch.log(classification) +
(1.0 - targets) * torch.log(1.0 - classification))
cls_loss = focal_weight * bce
zeros = torch.zeros_like(cls_loss)
if torch.cuda.is_available():
zeros = zeros.cuda()
cls_loss = torch.where(torch.ne(targets, -1.0), cls_loss, zeros)
classification_losses.append(
cls_loss.sum() /
torch.clamp(num_positive_anchors.to(dtype), min=1.0))
# rotation regression loss begin
if rotation_positive_indices.sum() > 0:
rotation_assigned_annotations_gt = hor_positive_assigned_annotations[
rotation_positive_indices]
rotation_anchor_widths_pi = anchor_widths[positive_index_list]
rotation_anchor_heights_pi = anchor_heights[
positive_index_list]
rotation_anchor_ctr_x_pi = anchor_ctr_x[positive_index_list]
rotation_anchor_ctr_y_pi = anchor_ctr_y[positive_index_list]
rotation_anchor_theta = anchor_theta[positive_index_list]
# efficientdet style
rotation_assigned_annotations_gt[:, 2] = torch.clamp(
rotation_assigned_annotations_gt[:, 2], min=1)
rotation_assigned_annotations_gt[:, 3] = torch.clamp(
rotation_assigned_annotations_gt[:, 3], min=1)
targets_dx = (
rotation_assigned_annotations_gt[:, 0] -
rotation_anchor_ctr_x_pi) / rotation_anchor_widths_pi
targets_dy = (
rotation_assigned_annotations_gt[:, 1] -
rotation_anchor_ctr_y_pi) / rotation_anchor_heights_pi
targets_dw = torch.log(rotation_assigned_annotations_gt[:, 2] /
rotation_anchor_widths_pi)
targets_dh = torch.log(rotation_assigned_annotations_gt[:, 3] /
rotation_anchor_heights_pi)
targets_theta = (
(rotation_assigned_annotations_gt[:, 4] / 180 * math.pi) -
(rotation_anchor_theta / 180 * math.pi))
targets = torch.stack((targets_dx, targets_dy, targets_dw,
targets_dh, targets_theta))
targets = targets.t()
regression_diff = torch.abs(targets -
regression[positive_index_list, :])
# smooth l1 loss with beta
# regression_loss = torch.where(
# torch.le(regression_diff, 1.0 / 9.0),
# 0.5 * 9.0 * torch.pow(regression_diff, 2),
# regression_diff - 0.5 / 9.0
# )
# smooth l1 loss
regression_loss = torch.where(
torch.le(regression_diff, 1),
0.5 * torch.pow(regression_diff, 2), regression_diff - 0.5)
regression_losses.append(regression_loss.mean())
else:
if torch.cuda.is_available():
regression_losses.append(torch.tensor(0).to(dtype).cuda())
else:
regression_losses.append(torch.tensor(0).to(dtype))
return torch.stack(classification_losses).mean(dim=0, keepdim=True), \
torch.stack(regression_losses).mean(dim=0,
keepdim=True) * 50