def eval(gt, det):
gt_bbox_num = 0
det_bbox_num = 0
tp = 0
miss_img = []
for img in gt.keys():
gt_bboxes = gt[img]
img_tp = 0
gt_bbox_num += len(gt_bboxes)
if img not in det:
continue
det_bbox_num += len(det[img])
for det_bbox in det[img]:
if (utils.iou_calc1(gt_bboxes, det_bbox[:4]) >= 0.5).any():
img_tp += 1
tp += img_tp
if(img_tp != len(det[img]) or img_tp != len(gt[img])):
miss_img.append(img)
recall = tp / gt_bbox_num
precision = tp / det_bbox_num
print('gt_bbox_num = %d' % gt_bbox_num)
print('det_bbox_num = %d' % det_bbox_num)
print('tp = %d' % tp)
print('F1 = %f' % (2. / (1. / recall + 1. / precision)))
return sorted(miss_img)
def dropObjectsInIgr(self, gt, det):
det = np.array(det)
idxFr = gt[:, 4] != 0
idxIgr = gt[:, 4] == 0
igrRegion = gt[idxIgr, 0:4]
igrDet = np.ones(len(det), dtype=np.bool)
for i in range(len(igrRegion)):
iou = iou_calc1(igrRegion[i], det[:, :4])
igrDet[iou > 0.5] = False
return gt[idxFr][:, [0, 1, 2, 3, 5]], det[igrDet]
def make_chip(self, sample, imgset):
# get image and mask informations
image = cv2.imread(sample['image'])
height, width = sample['height'], sample['width']
img_id = osp.splitext(osp.basename(sample['image']))[0]
mask_path = osp.join(self.segmentation_dir, '{}.hdf5'.format(img_id))
with h5py.File(mask_path, 'r') as hf:
mask = np.array(hf['label'])
mask_h, mask_w = mask.shape[:2]
# make chip
region_box, contours = utils.generate_box_from_mask(mask)
region_box = utils.generate_crop_region(region_box, mask,
(mask_w, mask_h),
(width, height), self.gbm,
args.aim)
region_box = utils.resize_box(region_box, (mask_w, mask_h),
(width, height))
# make tiling
if args.tiling and imgset != "val":
tiling = utils.add_tiling((width, height))
for pattern in tiling:
if utils.iou_calc1(pattern, region_box).max() < 0.85:
region_box = np.vstack((region_box, tiling))
if args.show:
utils.show_image(image[..., ::-1], np.array(region_box))
# get box and class
gt_bboxes, gt_cls = sample['bboxes'], sample['cls']
# generate chip annotations and writer chip image
chip_gt_list, chip_label_list, neglect_list = self.generate_region_gt(
region_box, gt_bboxes, gt_cls)
chip_loc = self.write_chip_and_anno(image, img_id, region_box,
chip_gt_list, chip_label_list,
neglect_list, imgset)
return chip_loc
def augPaster(self, chip_img, loc, bbox, labels):
# 当前chip中目标尺度和中心点, 寻找hardlabel
obj_scales = bbox[:, 2:4] - bbox[:, :2]
obj_center = (bbox[:, 2:4] + bbox[:, :2]) / 2
# hardsIn = [hard for hard in self.hard_cls if hard in labels]
# 抠出道路掩码并且创建粘贴点候选序列
road_chip = self.roadMask[loc[1]:loc[3], loc[0]:loc[2]].copy()
# utils.show_image(road_chip)
cand_chip = cv2.resize(road_chip, (0, 0),
fx=hpy["fx"],
fy=hpy["fy"],
interpolation=cv2.INTER_NEAREST)
cand_chip = cv2.erode(cand_chip, self.kernel)
# utils.show_image(cand_chip)
cand_points = np.argwhere(cand_chip[..., 0] >= 200) * (
road_chip.shape[0] / cand_chip.shape[0])
np.random.shuffle(cand_points)
# 粘贴
acount = 0 # 当前粘贴的次数
for time, point in enumerate(cand_points):
if time > 500: # 如果找了500次仍然没有找到合适的位置, 放弃
break
adj_obj_idx = np.argmin(np.power(point - obj_center, 2).sum(1))
adj_scale = obj_scales[adj_obj_idx]
adj_label = labels[adj_obj_idx]
adj_rank = self.scale_rank[adj_label]
# if acount < len(hardsIn):
# pasting_cls = hardsIn[acount]
# elif adj_label in self.aid_cls:
if adj_label in self.aid_cls:
pasting_cls = np.random.choice(self.aid_cls[adj_label])
else:
continue
scale_map = self.scale_rank[pasting_cls] / adj_rank
scale = np.array(adj_scale * scale_map)
# 寻找粘贴对象
paster_cls_pool = np.array(self.paster_pool[pasting_cls])
paster_shape = np.array(paster_cls_pool[:, [4, 5, 6]],
dtype=np.float)
mask1 = paster_shape[:,
scale.argmax()] > paster_shape[:,
scale.argmin()]
mask2 = paster_shape[:, scale.argmax()] >= 0.75 * scale.max()
mask3 = paster_shape[:, scale.argmax()] <= 2.5 * scale.max()
# mask4 = paster_shape[:, 2] >= 0.5 * (scale[0] / scale[1])
# mask5 = paster_shape[:, 2] <= 1.5 * (scale[0] / scale[1])
mask = mask1 & mask2 & mask3
paster_cls_pool = paster_cls_pool[mask]
if len(paster_cls_pool) == 0:
continue
paster_idx = np.random.choice(np.arange(len(paster_cls_pool)))
paster_info = paster_cls_pool[paster_idx]
paster_path = osp.join(self.maskPools_dir, "_".join(paster_info))
paster = cv2.imread(paster_path)
# 获取粘贴位置, 判断位置是否合法
paster_box = self.getPastingLocal(chip_img, paster, scale, point)
paster_area = np.product(paster_box[2:] - paster_box[:2])
paster_road_area = road_chip[paster_box[1]:paster_box[3],
paster_box[0]:paster_box[2]].sum()
if utils.iou_calc1(paster_box, bbox).max(
) > hpy["obt"] or paster_road_area < paster_area * hpy["ort"]:
continue
# 调整明亮度, 粘贴
paster = utils.adjustLumin(chip_img, paster,
float(paster_info[-2]), hpy['alpha'])
chip_img = self.pasting(chip_img, paster, paster_box)
# utils.show_image(chip_img, np.array([paster_box]))
# 添加尺度和中心信息、加入chip标签
obj_center = np.vstack(
(obj_center, paster_box[2:4] + paster_box[:2] / 2))
obj_scales = np.vstack(
(obj_scales, paster_box[2:4] - paster_box[:2]))
labels = np.hstack((labels, int(pasting_cls)))
bbox = np.vstack((bbox, paster_box))
# 计数
self.paster_num[int(pasting_cls)] += 1
acount += 1
if acount >= hpy["pasting_maximum"]:
break
return chip_img, bbox, labels