| 375 | self.dpg_prob = dpg_prob |
| 376 | |
| 377 | def __call__(self, results): |
| 378 | if np.random.rand() > self.dpg_prob: |
| 379 | return results |
| 380 | |
| 381 | bbox = results['bbox'] |
| 382 | imgwidth = results['ann_info']['width'] |
| 383 | imgheight = results['ann_info']['height'] |
| 384 | |
| 385 | PatchScale = random.uniform(0, 1) |
| 386 | width = bbox[2] - bbox[0] |
| 387 | ht = bbox[3] - bbox[1] |
| 388 | |
| 389 | if PatchScale > 0.85: |
| 390 | ratio = ht / width |
| 391 | if (width < ht): |
| 392 | patchWidth = PatchScale * width |
| 393 | patchHt = patchWidth * ratio |
| 394 | else: |
| 395 | patchHt = PatchScale * ht |
| 396 | patchWidth = patchHt / ratio |
| 397 | |
| 398 | xmin = bbox[0] + random.uniform(0, 1) * (width - patchWidth) |
| 399 | ymin = bbox[1] + random.uniform(0, 1) * (ht - patchHt) |
| 400 | xmax = xmin + patchWidth + 1 |
| 401 | ymax = ymin + patchHt + 1 |
| 402 | else: |
| 403 | xmin = max( |
| 404 | 1, |
| 405 | min(bbox[0] + np.random.normal(-0.0142, 0.1158) * width, |
| 406 | imgwidth - 3)) |
| 407 | ymin = max( |
| 408 | 1, |
| 409 | min(bbox[1] + np.random.normal(0.0043, 0.068) * ht, |
| 410 | imgheight - 3)) |
| 411 | xmax = min( |
| 412 | max(xmin + 2, |
| 413 | bbox[2] + np.random.normal(0.0154, 0.1337) * width), |
| 414 | imgwidth - 3) |
| 415 | ymax = min( |
| 416 | max(ymin + 2, |
| 417 | bbox[3] + np.random.normal(-0.0013, 0.0711) * ht), |
| 418 | imgheight - 3) |
| 419 | bbox_xyxy = np.array([xmin, ymin, xmax, ymax]) |
| 420 | bbox_xywh = xyxy2xywh(bbox_xyxy) |
| 421 | center, scale = box2cs(bbox_xywh, |
| 422 | aspect_ratio=1.0, |
| 423 | bbox_scale_factor=1.0) |
| 424 | results['bbox'] = bbox_xyxy |
| 425 | results['center'] = center |
| 426 | results['scale'] = scale |
| 427 | |
| 428 | return results |
| 429 | |
| 430 | |
| 431 | @PIPELINES.register_module() |