def preprocess(self, image):
height, width = image.shape[0:2]
# 使用模型输入分辨率还是保持原图分辨率
if not self.cfg.keep_res:
inp_height, inp_width = self.cfg.input_h, self.cfg.input_w
c = np.array([width / 2., height / 2.], dtype=np.float32)
s = max(height, width) * 1.0
else:
# 保证分辨率能被pad+1整除
inp_height = (height | self.cfg.pad) + 1
inp_width = (width | self.cfg.pad) + 1
c = np.array([width // 2, height // 2], dtype=np.float32)
s = np.array([inp_width, inp_height], dtype=np.float32)
trans_input = get_affine_transform(c, s, 0, [inp_width, inp_height])
inp_image = cv2.warpAffine(image,
trans_input, (inp_width, inp_height),
flags=cv2.INTER_LINEAR)
inp_image = ((inp_image / 255. - self.mean) / self.std).astype(
np.float32)
images = inp_image.transpose(2, 0, 1).reshape(1, 3, inp_height,
inp_width)
images = torch.from_numpy(images)
# 记录变换参数,用于检测结果反变换
meta = {
'c': c,
's': s,
'out_height': inp_height // self.cfg.down_ratio,
'out_width': inp_width // self.cfg.down_ratio
}
return images, meta
def pre_process(self, image, scale, calib=None):
height, width = image.shape[0:2]
inp_height, inp_width = self.opt.input_h, self.opt.input_w
c = np.array([width / 2, height / 2], dtype=np.float32)
if self.opt.keep_res:
s = np.array([inp_width, inp_height], dtype=np.int32)
else:
s = np.array([width, height], dtype=np.int32)
trans_input = get_affine_transform(c, s, 0, [inp_width, inp_height])
resized_image = image #cv2.resize(image, (width, height))
inp_image = cv2.warpAffine(resized_image,
trans_input, (inp_width, inp_height),
flags=cv2.INTER_LINEAR)
inp_image = (inp_image.astype(np.float32) / 255.)
inp_image = (inp_image - self.mean) / self.std
images = inp_image.transpose(2, 0, 1)[np.newaxis, ...]
calib = np.array(calib, dtype=np.float32) if calib is not None \
else self.calib
images = torch.from_numpy(images)
meta = {
'c': c,
's': s,
'out_height': inp_height // self.opt.down_ratio,
'out_width': inp_width // self.opt.down_ratio,
'calib': calib
}
return images, meta
def pre_process(self, image, scale, meta=None):
height, width = image.shape[0:2]
new_height = int(height * scale)
new_width = int(width * scale)
if self.opt.fix_res:
inp_height, inp_width = self.opt.input_h, self.opt.input_w
c = np.array([new_width / 2., new_height / 2.], dtype=np.float32)
s = max(height, width) * 1.0
else:
inp_height = (new_height | self.opt.pad) + 1
inp_width = (new_width | self.opt.pad) + 1
c = np.array([new_width // 2, new_height // 2], dtype=np.float32)
s = np.array([inp_width, inp_height], dtype=np.float32)
trans_input = get_affine_transform(c, s, 0, [inp_width, inp_height])
resized_image = cv2.resize(image, (new_width, new_height))
inp_image = cv2.warpAffine(
resized_image, trans_input, (inp_width, inp_height),
flags=cv2.INTER_LINEAR)
inp_image = ((inp_image / 255. - self.mean) / self.std).astype(np.float32)
images = inp_image.transpose(2, 0, 1).reshape(1, 3, inp_height, inp_width)
if self.opt.flip_test:
images = np.concatenate((images, images[:, :, :, ::-1]), axis=0)
images = torch.from_numpy(images)
meta = {'c': c, 's': s,
'out_height': inp_height // self.opt.down_ratio,
'out_width': inp_width // self.opt.down_ratio}
return images, meta
def __getitem__(self, index):
img_id = self.images[index]
file_name = self.coco.loadImgs(ids=[img_id])[0]['file_name']
img_path = os.path.join(self.img_dir, file_name)
ann_ids = self.coco.getAnnIds(imgIds=[img_id])
anns = self.coco.loadAnns(ids=ann_ids)
num_objs = min(len(anns), self.max_objs)
img = cv2.imread(img_path)
height, width = img.shape[0], img.shape[1]
c = np.array([img.shape[1] / 2., img.shape[0] / 2.], dtype=np.float32)
if self.opt.keep_res:
input_h = (height | self.opt.pad) + 1
input_w = (width | self.opt.pad) + 1
s = np.array([input_w, input_h], dtype=np.float32)
else:
s = max(img.shape[0], img.shape[1]) * 1.0
input_h, input_w = self.opt.input_h, self.opt.input_w
flipped = False
if self.split == 'train':
if not self.opt.not_rand_crop:
s = s * np.random.choice(np.arange(0.6, 1.4, 0.1))
w_border = self._get_border(128, img.shape[1])
h_border = self._get_border(128, img.shape[0])
c[0] = np.random.randint(low=w_border,
high=img.shape[1] - w_border)
c[1] = np.random.randint(low=h_border,
high=img.shape[0] - h_border)
else:
sf = self.opt.scale
cf = self.opt.shift
c[0] += s * np.clip(np.random.randn() * cf, -2 * cf, 2 * cf)
c[1] += s * np.clip(np.random.randn() * cf, -2 * cf, 2 * cf)
s = s * np.clip(np.random.randn() * sf + 1, 1 - sf, 1 + sf)
if np.random.random() < self.opt.flip:
flipped = True
img = img[:, ::-1, :]
c[0] = width - c[0] - 1
trans_input = get_affine_transform(c, s, 0, [input_w, input_h])
inp = cv2.warpAffine(img,
trans_input, (input_w, input_h),
flags=cv2.INTER_LINEAR)
inp = (inp.astype(np.float32) / 255.)
if self.split == 'train' and not self.opt.no_color_aug:
color_aug(self._data_rng, inp, self._eig_val, self._eig_vec)
inp = (inp - self.mean) / self.std
inp = inp.transpose(2, 0, 1)
output_h = input_h // self.opt.down_ratio
output_w = input_w // self.opt.down_ratio
num_classes = self.num_classes
trans_output = get_affine_transform(c, s, 0, [output_w, output_h])
hm = np.zeros((num_classes, output_h, output_w), dtype=np.float32)
wh = np.zeros((self.max_objs, 2), dtype=np.float32)
dense_wh = np.zeros((2, output_h, output_w), dtype=np.float32)
reg = np.zeros((self.max_objs, 2), dtype=np.float32)
ind = np.zeros((self.max_objs), dtype=np.int64)
reg_mask = np.zeros((self.max_objs), dtype=np.uint8)
cat_spec_wh = np.zeros((self.max_objs, num_classes * 2),
dtype=np.float32)
cat_spec_mask = np.zeros((self.max_objs, num_classes * 2),
dtype=np.uint8)
draw_gaussian = draw_msra_gaussian if self.opt.mse_loss else \
draw_umich_gaussian
gt_det = []
for k in range(num_objs):
ann = anns[k]
bbox = self._coco_box_to_bbox(ann['bbox'])
cls_id = int(self.cat_ids[ann['category_id']])
if flipped:
bbox[[0, 2]] = width - bbox[[2, 0]] - 1
bbox[:2] = affine_transform(bbox[:2], trans_output)
bbox[2:] = affine_transform(bbox[2:], trans_output)
bbox[[0, 2]] = np.clip(bbox[[0, 2]], 0, output_w - 1)
bbox[[1, 3]] = np.clip(bbox[[1, 3]], 0, output_h - 1)
h, w = bbox[3] - bbox[1], bbox[2] - bbox[0]
if h > 0 and w > 0:
radius = gaussian_radius((math.ceil(h), math.ceil(w)))
radius = max(0, int(radius))
radius = self.opt.hm_gauss if self.opt.mse_loss else radius
ct = np.array([(bbox[0] + bbox[2]) / 2,
(bbox[1] + bbox[3]) / 2],
dtype=np.float32)
ct_int = ct.astype(np.int32)
draw_gaussian(hm[cls_id], ct_int, radius)
wh[k] = 1. * w, 1. * h
ind[k] = ct_int[1] * output_w + ct_int[0]
reg[k] = ct - ct_int
reg_mask[k] = 1
cat_spec_wh[k, cls_id * 2:cls_id * 2 + 2] = wh[k]
cat_spec_mask[k, cls_id * 2:cls_id * 2 + 2] = 1
if self.opt.dense_wh:
draw_dense_reg(dense_wh, hm.max(axis=0), ct_int, wh[k],
radius)
gt_det.append([
ct[0] - w / 2, ct[1] - h / 2, ct[0] + w / 2, ct[1] + h / 2,
1, cls_id
])
ret = {
'input': inp,
'hm': hm,
'reg_mask': reg_mask,
'ind': ind,
'wh': wh
}
if self.opt.dense_wh:
hm_a = hm.max(axis=0, keepdims=True)
dense_wh_mask = np.concatenate([hm_a, hm_a], axis=0)
ret.update({'dense_wh': dense_wh, 'dense_wh_mask': dense_wh_mask})
del ret['wh']
elif self.opt.cat_spec_wh:
ret.update({
'cat_spec_wh': cat_spec_wh,
'cat_spec_mask': cat_spec_mask
})
del ret['wh']
if self.opt.reg_offset:
ret.update({'reg': reg})
if self.opt.debug > 0 or not self.split == 'train':
gt_det = np.array(gt_det, dtype=np.float32) if len(gt_det) > 0 else \
np.zeros((1, 6), dtype=np.float32)
meta = {'c': c, 's': s, 'gt_det': gt_det, 'img_id': img_id}
ret['meta'] = meta
return ret
def __getitem__(self, index):
img_id = self.images[index]
img_info = self.coco.loadImgs(ids=[img_id])[0]
img_path = os.path.join(self.img_dir, img_info['file_name'])
img = cv2.imread(img_path)
if 'calib' in img_info:
calib = np.array(img_info['calib'], dtype=np.float32)
else:
calib = self.calib
height, width = img.shape[0], img.shape[1]
c = np.array([img.shape[1] / 2., img.shape[0] / 2.])
if self.opt.keep_res:
s = np.array([self.opt.input_w, self.opt.input_h], dtype=np.int32)
else:
s = np.array([width, height], dtype=np.int32)
aug = False
if self.split == 'train' and np.random.random() < self.opt.aug_ddd:
aug = True
sf = self.opt.scale
cf = self.opt.shift
s = s * np.clip(np.random.randn() * sf + 1, 1 - sf, 1 + sf)
c[0] += img.shape[1] * np.clip(np.random.randn() * cf, -2 * cf,
2 * cf)
c[1] += img.shape[0] * np.clip(np.random.randn() * cf, -2 * cf,
2 * cf)
trans_input = get_affine_transform(
c, s, 0, [self.opt.input_w, self.opt.input_h])
inp = cv2.warpAffine(img,
trans_input, (self.opt.input_w, self.opt.input_h),
flags=cv2.INTER_LINEAR)
inp = (inp.astype(np.float32) / 255.)
# if self.split == 'train' and not self.opt.no_color_aug:
# color_aug(self._data_rng, inp, self._eig_val, self._eig_vec)
inp = (inp - self.mean) / self.std
inp = inp.transpose(2, 0, 1)
num_classes = self.opt.num_classes
trans_output = get_affine_transform(
c, s, 0, [self.opt.output_w, self.opt.output_h])
hm = np.zeros((num_classes, self.opt.output_h, self.opt.output_w),
dtype=np.float32)
wh = np.zeros((self.max_objs, 2), dtype=np.float32)
reg = np.zeros((self.max_objs, 2), dtype=np.float32)
dep = np.zeros((self.max_objs, 1), dtype=np.float32)
rotbin = np.zeros((self.max_objs, 2), dtype=np.int64)
rotres = np.zeros((self.max_objs, 2), dtype=np.float32)
dim = np.zeros((self.max_objs, 3), dtype=np.float32)
ind = np.zeros((self.max_objs), dtype=np.int64)
reg_mask = np.zeros((self.max_objs), dtype=np.uint8)
rot_mask = np.zeros((self.max_objs), dtype=np.uint8)
ann_ids = self.coco.getAnnIds(imgIds=[img_id])
anns = self.coco.loadAnns(ids=ann_ids)
num_objs = min(len(anns), self.max_objs)
draw_gaussian = draw_msra_gaussian if self.opt.mse_loss else \
draw_umich_gaussian
gt_det = []
for k in range(num_objs):
ann = anns[k]
bbox = self._coco_box_to_bbox(ann['bbox'])
cls_id = int(self.cat_ids[ann['category_id']])
if cls_id <= -99:
continue
# if flipped:
# bbox[[0, 2]] = width - bbox[[2, 0]] - 1
bbox[:2] = affine_transform(bbox[:2], trans_output)
bbox[2:] = affine_transform(bbox[2:], trans_output)
bbox[[0, 2]] = np.clip(bbox[[0, 2]], 0, self.opt.output_w - 1)
bbox[[1, 3]] = np.clip(bbox[[1, 3]], 0, self.opt.output_h - 1)
h, w = bbox[3] - bbox[1], bbox[2] - bbox[0]
if h > 0 and w > 0:
radius = gaussian_radius((h, w))
radius = max(0, int(radius))
ct = np.array([(bbox[0] + bbox[2]) / 2,
(bbox[1] + bbox[3]) / 2],
dtype=np.float32)
ct_int = ct.astype(np.int32)
if cls_id < 0:
ignore_id = [_ for _ in range(num_classes)] \
if cls_id == - 1 else [- cls_id - 2]
if self.opt.rect_mask:
hm[ignore_id,
int(bbox[1]):int(bbox[3]) + 1,
int(bbox[0]):int(bbox[2]) + 1] = 0.9999
else:
for cc in ignore_id:
draw_gaussian(hm[cc], ct, radius)
hm[ignore_id, ct_int[1], ct_int[0]] = 0.9999
continue
draw_gaussian(hm[cls_id], ct, radius)
wh[k] = 1. * w, 1. * h
gt_det.append([ct[0], ct[1], 1] + \
self._alpha_to_8(self._convert_alpha(ann['alpha'])) + \
[ann['depth']] + (np.array(ann['dim']) / 1).tolist() + [cls_id])
if self.opt.reg_bbox:
gt_det[-1] = gt_det[-1][:-1] + [w, h] + [gt_det[-1][-1]]
# if (not self.opt.car_only) or cls_id == 1: # Only estimate ADD for cars !!!
if 1:
alpha = self._convert_alpha(ann['alpha'])
# print('img_id cls_id alpha rot_y', img_path, cls_id, alpha, ann['rotation_y'])
if alpha < np.pi / 6. or alpha > 5 * np.pi / 6.:
rotbin[k, 0] = 1
rotres[k, 0] = alpha - (-0.5 * np.pi)
if alpha > -np.pi / 6. or alpha < -5 * np.pi / 6.:
rotbin[k, 1] = 1
rotres[k, 1] = alpha - (0.5 * np.pi)
dep[k] = ann['depth']
dim[k] = ann['dim']
# print(' cat dim', cls_id, dim[k])
ind[k] = ct_int[1] * self.opt.output_w + ct_int[0]
reg[k] = ct - ct_int
reg_mask[k] = 1 if not aug else 0
rot_mask[k] = 1
# print('gt_det', gt_det)
# print('')
ret = {
'input': inp,
'hm': hm,
'dep': dep,
'dim': dim,
'ind': ind,
'rotbin': rotbin,
'rotres': rotres,
'reg_mask': reg_mask,
'rot_mask': rot_mask
}
if self.opt.reg_bbox:
ret.update({'wh': wh})
if self.opt.reg_offset:
ret.update({'reg': reg})
if self.opt.debug > 0 or not ('train' in self.split):
gt_det = np.array(gt_det, dtype=np.float32) if len(gt_det) > 0 else \
np.zeros((1, 18), dtype=np.float32)
meta = {
'c': c,
's': s,
'gt_det': gt_det,
'calib': calib,
'image_path': img_path,
'img_id': img_id
}
ret['meta'] = meta
return ret
def __getitem__(self, index):
img_id = self.images[index]
file_name = self.coco.loadImgs(ids=[img_id])[0]['file_name']
img_path = os.path.join(self.img_dir, file_name)
ann_ids = self.coco.getAnnIds(imgIds=[img_id])
anns = self.coco.loadAnns(ids=ann_ids)
num_objs = min(len(anns), self.max_objs)
img = cv2.imread(img_path)
height, width = img.shape[0], img.shape[1]
c = np.array([img.shape[1] / 2., img.shape[0] / 2.], dtype=np.float32)
s = max(img.shape[0], img.shape[1]) * 1.0
rot = 0
flipped = False
if self.split == 'train':
if not self.opt.not_rand_crop:
s = s * np.random.choice(np.arange(0.6, 1.4, 0.1))
w_border = self._get_border(128, img.shape[1])
h_border = self._get_border(128, img.shape[0])
c[0] = np.random.randint(low=w_border,
high=img.shape[1] - w_border)
c[1] = np.random.randint(low=h_border,
high=img.shape[0] - h_border)
else:
sf = self.opt.scale
cf = self.opt.shift
c[0] += s * np.clip(np.random.randn() * cf, -2 * cf, 2 * cf)
c[1] += s * np.clip(np.random.randn() * cf, -2 * cf, 2 * cf)
s = s * np.clip(np.random.randn() * sf + 1, 1 - sf, 1 + sf)
if np.random.random() < self.opt.aug_rot:
rf = self.opt.rotate
rot = np.clip(np.random.randn() * rf, -rf * 2, rf * 2)
if np.random.random() < self.opt.flip:
flipped = True
img = img[:, ::-1, :]
c[0] = width - c[0] - 1
trans_input = get_affine_transform(
c, s, rot, [self.opt.input_res, self.opt.input_res])
inp = cv2.warpAffine(img,
trans_input,
(self.opt.input_res, self.opt.input_res),
flags=cv2.INTER_LINEAR)
inp = (inp.astype(np.float32) / 255.)
if self.split == 'train' and not self.opt.no_color_aug:
color_aug(self._data_rng, inp, self._eig_val, self._eig_vec)
inp = (inp - self.mean) / self.std
inp = inp.transpose(2, 0, 1)
output_res = self.opt.output_res
num_joints = self.num_joints
trans_output_rot = get_affine_transform(c, s, rot,
[output_res, output_res])
trans_output = get_affine_transform(c, s, 0, [output_res, output_res])
hm = np.zeros((self.num_classes, output_res, output_res),
dtype=np.float32)
hm_hp = np.zeros((num_joints, output_res, output_res),
dtype=np.float32)
dense_kps = np.zeros((num_joints, 2, output_res, output_res),
dtype=np.float32)
dense_kps_mask = np.zeros((num_joints, output_res, output_res),
dtype=np.float32)
wh = np.zeros((self.max_objs, 2), dtype=np.float32)
kps = np.zeros((self.max_objs, num_joints * 2), dtype=np.float32)
reg = np.zeros((self.max_objs, 2), dtype=np.float32)
ind = np.zeros((self.max_objs), dtype=np.int64)
reg_mask = np.zeros((self.max_objs), dtype=np.uint8)
kps_mask = np.zeros((self.max_objs, self.num_joints * 2),
dtype=np.uint8)
hp_offset = np.zeros((self.max_objs * num_joints, 2), dtype=np.float32)
hp_ind = np.zeros((self.max_objs * num_joints), dtype=np.int64)
hp_mask = np.zeros((self.max_objs * num_joints), dtype=np.int64)
draw_gaussian = draw_msra_gaussian if self.opt.mse_loss else \
draw_umich_gaussian
gt_det = []
for k in range(num_objs):
ann = anns[k]
bbox = self._coco_box_to_bbox(ann['bbox'])
cls_id = int(ann['category_id']) - 1
pts = np.array(ann['keypoints'], np.float32).reshape(num_joints, 3)
if flipped:
bbox[[0, 2]] = width - bbox[[2, 0]] - 1
pts[:, 0] = width - pts[:, 0] - 1
for e in self.flip_idx:
pts[e[0]], pts[e[1]] = pts[e[1]].copy(), pts[e[0]].copy()
bbox[:2] = affine_transform(bbox[:2], trans_output)
bbox[2:] = affine_transform(bbox[2:], trans_output)
bbox = np.clip(bbox, 0, output_res - 1)
h, w = bbox[3] - bbox[1], bbox[2] - bbox[0]
if (h > 0 and w > 0) or (rot != 0):
radius = gaussian_radius((math.ceil(h), math.ceil(w)))
radius = self.opt.hm_gauss if self.opt.mse_loss else max(
0, int(radius))
ct = np.array([(bbox[0] + bbox[2]) / 2,
(bbox[1] + bbox[3]) / 2],
dtype=np.float32)
ct_int = ct.astype(np.int32)
wh[k] = 1. * w, 1. * h
ind[k] = ct_int[1] * output_res + ct_int[0]
reg[k] = ct - ct_int
reg_mask[k] = 1
num_kpts = pts[:, 2].sum()
if num_kpts == 0:
hm[cls_id, ct_int[1], ct_int[0]] = 0.9999
reg_mask[k] = 0
hp_radius = gaussian_radius((math.ceil(h), math.ceil(w)))
hp_radius = self.opt.hm_gauss \
if self.opt.mse_loss else max(0, int(hp_radius))
for j in range(num_joints):
if pts[j, 2] > 0:
pts[j, :2] = affine_transform(pts[j, :2],
trans_output_rot)
if pts[j, 0] >= 0 and pts[j, 0] < output_res and \
pts[j, 1] >= 0 and pts[j, 1] < output_res:
kps[k, j * 2:j * 2 + 2] = pts[j, :2] - ct_int
kps_mask[k, j * 2:j * 2 + 2] = 1
pt_int = pts[j, :2].astype(np.int32)
hp_offset[k * num_joints + j] = pts[j, :2] - pt_int
hp_ind[k * num_joints +
j] = pt_int[1] * output_res + pt_int[0]
hp_mask[k * num_joints + j] = 1
if self.opt.dense_hp:
# must be before draw center hm gaussian
draw_dense_reg(dense_kps[j],
hm[cls_id],
ct_int,
pts[j, :2] - ct_int,
radius,
is_offset=True)
draw_gaussian(dense_kps_mask[j], ct_int,
radius)
draw_gaussian(hm_hp[j], pt_int, hp_radius)
draw_gaussian(hm[cls_id], ct_int, radius)
gt_det.append([
ct[0] - w / 2, ct[1] - h / 2, ct[0] + w / 2, ct[1] +
h / 2, 1
] + pts[:, :2].reshape(num_joints * 2).tolist() + [cls_id])
if rot != 0:
hm = hm * 0 + 0.9999
reg_mask *= 0
kps_mask *= 0
ret = {
'input': inp,
'hm': hm,
'reg_mask': reg_mask,
'ind': ind,
'wh': wh,
'hps': kps,
'hps_mask': kps_mask
}
if self.opt.dense_hp:
dense_kps = dense_kps.reshape(num_joints * 2, output_res,
output_res)
dense_kps_mask = dense_kps_mask.reshape(num_joints, 1, output_res,
output_res)
dense_kps_mask = np.concatenate([dense_kps_mask, dense_kps_mask],
axis=1)
dense_kps_mask = dense_kps_mask.reshape(num_joints * 2, output_res,
output_res)
ret.update({
'dense_hps': dense_kps,
'dense_hps_mask': dense_kps_mask
})
del ret['hps'], ret['hps_mask']
if self.opt.reg_offset:
ret.update({'reg': reg})
if self.opt.hm_hp:
ret.update({'hm_hp': hm_hp})
if self.opt.reg_hp_offset:
ret.update({
'hp_offset': hp_offset,
'hp_ind': hp_ind,
'hp_mask': hp_mask
})
if self.opt.debug > 0 or not self.split == 'train':
gt_det = np.array(gt_det, dtype=np.float32) if len(gt_det) > 0 else \
np.zeros((1, 40), dtype=np.float32)
meta = {'c': c, 's': s, 'gt_det': gt_det, 'img_id': img_id}
ret['meta'] = meta
return ret
def __getitem__(self, index):
img_id = self.images[index]
file_name = self.coco.loadImgs(ids=[img_id])[0]['file_name']
img_path = os.path.join(self.img_dir, file_name)
# print("55555:", img_path)
ann_ids = self.coco.getAnnIds(
imgIds=[img_id]) # getAnnIds:通过输入图片的id来得到图片的anno_id
anns = self.coco.loadAnns(
ids=ann_ids) # loadAnns:通过anno_id，得到图片对应的详细anno信息
# print(111111, anns)
num_objs = min(len(anns), self.max_objs)
img = cv2.imread(img_path)
height, width = img.shape[0], img.shape[1]
c = np.array([img.shape[1] / 2., img.shape[0] / 2.], dtype=np.float32)
if self.opt.keep_res:
input_h = (height | self.opt.pad) + 1
input_w = (width | self.opt.pad) + 1
s = np.array([input_w, input_h], dtype=np.float32)
else:
s = max(img.shape[0], img.shape[1]) * 1.0
input_h, input_w = self.opt.input_h, self.opt.input_w
flipped = False
if self.split == 'train':
if not self.opt.not_rand_crop: # (not self.opt.not_rand_crop) = True
# s = s * np.random.choice(np.arange(0.6, 1.4, 0.1)) # 从0.6-1.4中随机选取一个数字（步长为0.1）
s = s * np.random.choice(np.arange(
0.7, 1.3, 0.1)) # 从0.6-1.4中随机选取一个数字（步长为0.1）
w_border = self._get_border(128, img.shape[1])
h_border = self._get_border(128, img.shape[0])
c[0] = np.random.randint(low=w_border,
high=img.shape[1] - w_border)
c[1] = np.random.randint(low=h_border,
high=img.shape[0] - h_border)
# print(111, c)
else:
sf = self.opt.scale # 0.4 # when not using random crop apply scale augmentation.
cf = self.opt.shift # 0.1 # when not using random crop apply shift augmentation.
c[0] += s * np.clip(np.random.randn() * cf, -2 * cf, 2 * cf)
c[1] += s * np.clip(np.random.randn() * cf, -2 * cf, 2 * cf)
s = s * np.clip(np.random.randn() * sf + 1, 1 - sf, 1 + sf)
if np.random.random() < self.opt.flip:
flipped = True
img = img[:, ::-1, :]
c[0] = width - c[0] - 1
trans_input = get_affine_transform(c, s, 0,
[input_w, input_h]) # 由三对点计算仿射变换
inp = cv2.warpAffine(img,
trans_input, (input_w, input_h),
flags=cv2.INTER_LINEAR) # 对图像做仿射变换
inp = (inp.astype(np.float32) / 255.) # 归一化
if self.split == 'train' and not self.opt.no_color_aug:
color_aug(self._data_rng, inp, self._eig_val, self._eig_vec)
# inp = (inp - self.mean) / self.std # ！！！！！！！ modified by zy
inp = inp.transpose(2, 0, 1)
# add by zy
# inp = np.transpose(inp, [1, 2, 0])
# cv2.imshow('input', inp)
# cv2.waitKey(0)
output_h = input_h // self.opt.down_ratio # 网络输出的预测结果特征图是128*128,这里要将GTbox也缩小为128*128来计算loss
output_w = input_w // self.opt.down_ratio #
num_classes = self.num_classes
trans_output = get_affine_transform(c, s, 0, [output_w, output_h])
hm = np.zeros((num_classes, output_h, output_w), dtype=np.float32)
wh = np.zeros((self.max_objs, 2), dtype=np.float32)
ori_wh = np.zeros((self.max_objs, 2), dtype=np.float32)
cxcy = np.zeros((self.max_objs, 2), dtype=np.float32)
ori_cxcy = np.zeros((self.max_objs, 2), dtype=np.float32)
cls_idx = np.zeros((self.max_objs), dtype=np.int64)
dense_wh = np.zeros((2, output_h, output_w), dtype=np.float32)
reg = np.zeros((self.max_objs, 2), dtype=np.float32)
ind = np.zeros((self.max_objs), dtype=np.int64)
reg_mask = np.zeros((self.max_objs), dtype=np.uint8)
cat_spec_wh = np.zeros((self.max_objs, num_classes * 2),
dtype=np.float32)
cat_spec_mask = np.zeros((self.max_objs, num_classes * 2),
dtype=np.uint8)
draw_gaussian = draw_msra_gaussian if self.opt.mse_loss else \
draw_umich_gaussian
gt_det = []
for k in range(num_objs):
ann = anns[k]
bbox = self._coco_box_to_bbox(ann['bbox'])
ori_h, ori_w = bbox[3] - bbox[1], bbox[2] - bbox[0]
ori_cx, ori_cy = (bbox[0] + bbox[2]) / 2, (bbox[1] + bbox[3]) / 2
cls_id = int(self.cat_ids[ann['category_id']])
if flipped:
bbox[[0, 2]] = width - bbox[[2, 0]] - 1
# print(111111, bbox)
bbox[:2] = affine_transform(bbox[:2], trans_output)
bbox[2:] = affine_transform(bbox[2:], trans_output)
# print(222222, bbox)
bbox[[0, 2]] = np.clip(bbox[[0, 2]], 0, output_w - 1)
bbox[[1, 3]] = np.clip(bbox[[1, 3]], 0, output_h - 1)
h, w = bbox[3] - bbox[1], bbox[2] - bbox[0]
cx, cy = (bbox[0] + bbox[2]) / 2, (bbox[1] + bbox[3]) / 2
# print(333333, h,w)
if h > 0 and w > 0:
radius = gaussian_radius((math.ceil(h), math.ceil(w)))
radius = max(0, int(radius))
radius = self.opt.hm_gauss if self.opt.mse_loss else radius
ct = np.array([(bbox[0] + bbox[2]) / 2,
(bbox[1] + bbox[3]) / 2],
dtype=np.float32)
ct_int = ct.astype(np.int32)
draw_gaussian(hm[cls_id], ct_int, radius)
wh[k] = 1. * w, 1. * h
ori_wh[k] = 1. * ori_w, 1. * ori_h
cxcy[k] = 1. * cx, 1. * cy
ori_cxcy[k] = 1. * ori_cx, 1. * ori_cy
cls_idx[k] = cls_id
ind[k] = ct_int[1] * output_w + ct_int[
0] # ind这个参数是用来？？？ ind = int_cy*output_w + int_cx
reg[k] = ct - ct_int # reg是用来回归精确小数与整数之间的误差的？？？
reg_mask[k] = 1
cat_spec_wh[k, cls_id * 2:cls_id * 2 + 2] = wh[k]
cat_spec_mask[k, cls_id * 2:cls_id * 2 + 2] = 1
if self.opt.dense_wh:
draw_dense_reg(dense_wh, hm.max(axis=0), ct_int, wh[k],
radius)
gt_det.append([
ct[0] - w / 2, ct[1] - h / 2, ct[0] + w / 2, ct[1] + h / 2,
1, cls_id
]) # [tlx,tly,brx,bry,1,cls_id]
ret = {
'input': inp,
'hm': hm,
'reg_mask': reg_mask,
'ind': ind,
'wh': wh,
'ori_wh': ori_wh,
'cxcy': cxcy,
'ori_cxcy': ori_cxcy,
'cls_idx': cls_idx
}
if self.opt.dense_wh: # FALSE , 'apply weighted regression near center or just apply regression on center point.'
hm_a = hm.max(axis=0, keepdims=True)
dense_wh_mask = np.concatenate([hm_a, hm_a], axis=0)
ret.update({'dense_wh': dense_wh, 'dense_wh_mask': dense_wh_mask})
del ret['wh']
elif self.opt.cat_spec_wh: # flase , 'category specific bounding box size.'
ret.update({
'cat_spec_wh': cat_spec_wh,
'cat_spec_mask': cat_spec_mask
})
del ret['wh']
if self.opt.reg_offset: # true
ret.update({'reg': reg})
if self.opt.debug > 0 or not self.split == 'train':
gt_det = np.array(gt_det, dtype=np.float32) if len(gt_det) > 0 else \
np.zeros((1, 6), dtype=np.float32)
meta = {'c': c, 's': s, 'gt_det': gt_det, 'img_id': img_id}
ret['meta'] = meta
# print(ret['hm'].shape, )
# cx, cy = int(ret['cxcy'][0][0]), int(ret['cxcy'][0][1])
# w, h = int(ret['wh'][0][0]), int(ret['wh'][0][1])
# print(cx,cy,w,h, ret['hm'][0][cy][cx])
# print(1111111111111, ret['input'].shape, ret['hm'].shape, ret['wh'].shape, ret['reg'].shape, ret['reg_mask'].shape, ret['ind'].shape)
# input_img = np.transpose([ret['hm'][cls_id]], [1,2,0])
# crop_img = input_img[(cy - h // 2):(cy + h // 2), (cx - w // 2):(cx + w // 2)]
# cv2.rectangle(input_img, (cx - w // 2, cy - h // 2),
# (cx + w // 2, cy + h // 2), (255, 250, 250), 2)
# cv2.imshow('input', crop_img)
# cv2.waitKey(0)
# print(2222222222222, type(ret['input']), type(ret['hm']), type(ret['hm']), type(ret['reg']), type(ret['reg_mask']), type(ret['ind']))
# print(ret.keys()) # dict_keys(['ind', 'reg_mask', 'hm', 'input', 'reg', 'wh'])
return ret
def __getitem__(self, index):
img_id = self.images[index]
img_info = self.coco.loadImgs(ids=[img_id])[0]
img_path = os.path.join(self.img_dir, img_info['file_name'])
img = cv2.imread(img_path)
height, width = img.shape[0], img.shape[1]
c = np.array([img.shape[1] / 2., img.shape[0] / 2.])
s = max(img.shape[0], img.shape[1]) * 1.0
flipped = False
if self.split == 'train':
if not self.opt.not_rand_crop:
s = s * np.random.choice(np.arange(0.6, 1.4, 0.1))
w_border = self._get_border(128, img.shape[1])
h_border = self._get_border(128, img.shape[0])
c[0] = np.random.randint(low=w_border,
high=img.shape[1] - w_border)
c[1] = np.random.randint(low=h_border,
high=img.shape[0] - h_border)
else:
sf = self.opt.scale
cf = self.opt.shift
s = s * np.clip(np.random.randn() * sf + 1, 1 - sf, 1 + sf)
c[0] += img.shape[1] * np.clip(np.random.randn() * cf, -2 * cf,
2 * cf)
c[1] += img.shape[0] * np.clip(np.random.randn() * cf, -2 * cf,
2 * cf)
if np.random.random() < self.opt.flip:
flipped = True
img = img[:, ::-1, :]
trans_input = get_affine_transform(
c, s, 0, [self.opt.input_res, self.opt.input_res])
inp = cv2.warpAffine(img,
trans_input,
(self.opt.input_res, self.opt.input_res),
flags=cv2.INTER_LINEAR)
inp = (inp.astype(np.float32) / 255.)
if self.split == 'train' and not self.opt.no_color_aug:
color_aug(self._data_rng, inp, self._eig_val, self._eig_vec)
inp = (inp - self.mean) / self.std
inp = inp.transpose(2, 0, 1)
output_res = self.opt.output_res
num_classes = self.opt.num_classes
trans_output = get_affine_transform(c, s, 0, [output_res, output_res])
num_hm = 1 if self.opt.agnostic_ex else num_classes
hm_t = np.zeros((num_hm, output_res, output_res), dtype=np.float32)
hm_l = np.zeros((num_hm, output_res, output_res), dtype=np.float32)
hm_b = np.zeros((num_hm, output_res, output_res), dtype=np.float32)
hm_r = np.zeros((num_hm, output_res, output_res), dtype=np.float32)
hm_c = np.zeros((num_classes, output_res, output_res),
dtype=np.float32)
reg_t = np.zeros((self.max_objs, 2), dtype=np.float32)
reg_l = np.zeros((self.max_objs, 2), dtype=np.float32)
reg_b = np.zeros((self.max_objs, 2), dtype=np.float32)
reg_r = np.zeros((self.max_objs, 2), dtype=np.float32)
ind_t = np.zeros((self.max_objs), dtype=np.int64)
ind_l = np.zeros((self.max_objs), dtype=np.int64)
ind_b = np.zeros((self.max_objs), dtype=np.int64)
ind_r = np.zeros((self.max_objs), dtype=np.int64)
reg_mask = np.zeros((self.max_objs), dtype=np.uint8)
ann_ids = self.coco.getAnnIds(imgIds=[img_id])
anns = self.coco.loadAnns(ids=ann_ids)
num_objs = min(len(anns), self.max_objs)
draw_gaussian = draw_msra_gaussian if self.opt.mse_loss else \
draw_umich_gaussian
for k in range(num_objs):
ann = anns[k]
# bbox = self._coco_box_to_bbox(ann['bbox'])
# tlbr
pts = np.array(ann['extreme_points'],
dtype=np.float32).reshape(4, 2)
# cls_id = int(self.cat_ids[ann['category_id']] - 1) # bug
cls_id = int(self.cat_ids[ann['category_id']])
hm_id = 0 if self.opt.agnostic_ex else cls_id
if flipped:
pts[:, 0] = width - pts[:, 0] - 1
pts[1], pts[3] = pts[3].copy(), pts[1].copy()
for j in range(4):
pts[j] = affine_transform(pts[j], trans_output)
pts = np.clip(pts, 0, self.opt.output_res - 1)
h, w = pts[2, 1] - pts[0, 1], pts[3, 0] - pts[1, 0]
if h > 0 and w > 0:
radius = gaussian_radius((math.ceil(h), math.ceil(w)))
radius = max(0, int(radius))
pt_int = pts.astype(np.int32)
draw_gaussian(hm_t[hm_id], pt_int[0], radius)
draw_gaussian(hm_l[hm_id], pt_int[1], radius)
draw_gaussian(hm_b[hm_id], pt_int[2], radius)
draw_gaussian(hm_r[hm_id], pt_int[3], radius)
reg_t[k] = pts[0] - pt_int[0]
reg_l[k] = pts[1] - pt_int[1]
reg_b[k] = pts[2] - pt_int[2]
reg_r[k] = pts[3] - pt_int[3]
ind_t[k] = pt_int[0, 1] * output_res + pt_int[0, 0]
ind_l[k] = pt_int[1, 1] * output_res + pt_int[1, 0]
ind_b[k] = pt_int[2, 1] * output_res + pt_int[2, 0]
ind_r[k] = pt_int[3, 1] * output_res + pt_int[3, 0]
ct = [
int((pts[3, 0] + pts[1, 0]) / 2),
int((pts[0, 1] + pts[2, 1]) / 2)
]
draw_gaussian(hm_c[cls_id], ct, radius)
reg_mask[k] = 1
ret = {
'input': inp,
'hm_t': hm_t,
'hm_l': hm_l,
'hm_b': hm_b,
'hm_r': hm_r,
'hm_c': hm_c
}
if self.opt.reg_offset:
ret.update({
'reg_mask': reg_mask,
'reg_t': reg_t,
'reg_l': reg_l,
'reg_b': reg_b,
'reg_r': reg_r,
'ind_t': ind_t,
'ind_l': ind_l,
'ind_b': ind_b,
'ind_r': ind_r
})
return ret
