| 190 | return imgori |
| 191 | |
| 192 | def resize_norm_img(self, data, gen_ratio, padding=True): |
| 193 | img = data['image'] |
| 194 | h = img.shape[0] |
| 195 | w = img.shape[1] |
| 196 | |
| 197 | imgW, imgH = self.base_shape[gen_ratio - 1] if gen_ratio <= 4 else [ |
| 198 | self.base_h * gen_ratio, self.base_h |
| 199 | ] |
| 200 | use_ratio = imgW // imgH |
| 201 | if use_ratio >= (w // h) + 2: |
| 202 | self.error += 1 |
| 203 | return None |
| 204 | if not padding: |
| 205 | resized_image = cv2.resize(img, (imgW, imgH), |
| 206 | interpolation=cv2.INTER_LINEAR) |
| 207 | resized_w = imgW |
| 208 | else: |
| 209 | ratio = w / float(h) |
| 210 | if math.ceil(imgH * ratio) > imgW: |
| 211 | resized_w = imgW |
| 212 | else: |
| 213 | resized_w = int( |
| 214 | math.ceil(imgH * ratio * (random.random() + 0.5))) |
| 215 | resized_w = min(imgW, resized_w) |
| 216 | |
| 217 | resized_image = cv2.resize(img, (resized_w, imgH)) |
| 218 | resized_image = resized_image.astype('float32') |
| 219 | resized_image = resized_image.transpose((2, 0, 1)) / 255 |
| 220 | resized_image -= 0.5 |
| 221 | resized_image /= 0.5 |
| 222 | padding_im = np.zeros((3, imgH, imgW), dtype=np.float32) |
| 223 | padding_im[:, :, :resized_w] = resized_image |
| 224 | valid_ratio = min(1.0, float(resized_w / imgW)) |
| 225 | data['image'] = padding_im |
| 226 | data['valid_ratio'] = valid_ratio |
| 227 | data['gen_ratio'] = imgW // imgH |
| 228 | data['real_ratio'] = max(1, round(w / h)) |
| 229 | return data |
| 230 | |
| 231 | def get_lmdb_sample_info(self, txn, index): |
| 232 | label_key = 'label-%09d'.encode() % index |