andevs commited on
Commit
cf81cff
·
verified ·
1 Parent(s): cb097e5

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +36 -27
app.py CHANGED
@@ -611,37 +611,41 @@ def _refine_video_alpha(result_rgba: Image.Image, guide_rgb: np.ndarray) -> Imag
611
 
612
  def _recover_video_mask_dropout(current: np.ndarray, raw: np.ndarray,
613
  previous: Optional[np.ndarray]) -> tuple:
614
- """Conservative, within-batch dropout recovery; never copies opaque old pixels.
615
 
616
- A previous mask is used only as a spatial hint for pixels which the *current*
617
- Silueta inference still assigns some foreground probability. This prevents
618
- replacing a moving subject with an old silhouette. No state survives batches.
619
  """
 
 
 
 
 
 
 
 
 
620
  if previous is None or previous.shape != current.shape:
621
  return current, False
622
- h, w = current.shape
623
  prev_area = int(np.count_nonzero(previous >= 128))
624
- cur_area = int(np.count_nonzero(current >= 128))
625
- if prev_area < h*w*0.025 or cur_area < h*w*0.008:
626
- return current, False
627
- ratio = cur_area / max(1, prev_area)
628
- if not (0.25 <= ratio < 0.63):
629
- return current, False
630
- prev_core = (previous >= 175).astype(np.uint8)
631
- cur_core = (current >= 175).astype(np.uint8)
632
- # Reject large subject motion; old outlines would cause ghosting.
633
- intersection = np.count_nonzero((prev_core > 0) & (cur_core > 0))
634
- if intersection / max(1, np.count_nonzero(cur_core)) < 0.65:
635
  return current, False
636
- # Recover only pixels already supported by current raw model confidence.
637
- near_previous = cv2.dilate(prev_core, np.ones((7, 7), np.uint8)) > 0
638
- candidate = near_previous & (raw >= 55) & (current < 128)
639
- if not np.any(candidate):
640
  return current, False
641
- out = current.copy()
642
- # A modest boost, never turn faint background evidence fully opaque.
643
- out[candidate] = np.maximum(out[candidate], np.minimum(raw[candidate], 175))
644
- return out, True
 
 
 
 
 
 
 
 
645
 
646
 
647
  def _reset_session(session_id: str) -> bool:
@@ -1538,8 +1542,11 @@ def _process_frames_sync(raw_frames, resolved_model, transparent,
1538
  arr[:, :, 3], raw_alpha, previous_alpha)
1539
  arr[:, :, 3] = recovered
1540
  result = Image.fromarray(arr, "RGBA")
1541
- # Store only reliable frames as references, never a recovered frame.
1542
- if not did_recover:
 
 
 
1543
  previous_alpha = recovered.copy()
1544
  blended = False
1545
  if active_session is not None:
@@ -1563,7 +1570,9 @@ def _process_frames_sync(raw_frames, resolved_model, transparent,
1563
  results.append({"data":base64.b64encode(buf.getvalue()).decode("ascii"),
1564
  "format":fmt,"size":buf.tell(),
1565
  "model":used_model,"smoothed":blended,
1566
- "matting":used_matting,"dropout_recovered":did_recover})
 
 
1567
  except Exception as e:
1568
  logger.error(f"[frames-batch] frame {idx} failed: {e}")
1569
  results.append({"data":"","format":"png","error":str(e)[:120]})
 
611
 
612
  def _recover_video_mask_dropout(current: np.ndarray, raw: np.ndarray,
613
  previous: Optional[np.ndarray]) -> tuple:
614
+ """Protect legitimate foreground against destructive refinement.
615
 
616
+ Never paste a previous frame into a moving subject: use only CURRENT-frame
617
+ Silueta confidence. Previous alpha is a diagnostic reference, not pixels
618
+ to composite. A truly missing raw mask cannot be reconstructed here.
619
  """
620
+ h, w = current.shape
621
+ raw_area = int(np.count_nonzero(raw >= 128))
622
+ refined_area = int(np.count_nonzero(current >= 128))
623
+ # The current model detected foreground that refinement threw away.
624
+ # Recover from this frame's original alpha, including disconnected limbs.
625
+ if raw_area > h * w * 0.015 and refined_area < raw_area * 0.78:
626
+ logger.warning('[video-mask] postprocess dropout: raw=%d refined=%d; restoring raw',
627
+ raw_area, refined_area)
628
+ return raw.copy(), True
629
  if previous is None or previous.shape != current.shape:
630
  return current, False
 
631
  prev_area = int(np.count_nonzero(previous >= 128))
632
+ if prev_area < h * w * 0.02 or raw_area < h * w * 0.008:
 
 
 
 
 
 
 
 
 
 
633
  return current, False
634
+ ratio = raw_area / max(prev_area, 1)
635
+ if ratio >= 0.62:
 
 
636
  return current, False
637
+ # Raw inference itself shrank. Previous mask must not be blindly copied;
638
+ # retaining current raw values avoids edge clipping, not model dropout.
639
+ prev_core = previous >= 160
640
+ cur_core = raw >= 160
641
+ overlap = np.count_nonzero(prev_core & cur_core) / max(1, np.count_nonzero(cur_core))
642
+ logger.warning('[video-mask] raw model dropout: previous=%d raw=%d ratio=%.3f overlap=%.3f',
643
+ prev_area, raw_area, ratio, overlap)
644
+ # Only relax cleanup if there is evidence of a shared, relatively static subject.
645
+ if overlap >= 0.70:
646
+ restored = np.maximum(current, np.where(raw >= 35, raw, 0).astype(np.uint8))
647
+ return restored, True
648
+ return current, False
649
 
650
 
651
  def _reset_session(session_id: str) -> bool:
 
1542
  arr[:, :, 3], raw_alpha, previous_alpha)
1543
  arr[:, :, 3] = recovered
1544
  result = Image.fromarray(arr, "RGBA")
1545
+ # Diagnostic metrics make the failure stage observable in server logs.
1546
+ raw_fraction = round(float(np.count_nonzero(raw_alpha >= 128)) / raw_alpha.size, 4)
1547
+ final_fraction = round(float(np.count_nonzero(recovered >= 128)) / recovered.size, 4)
1548
+ # Never use a suspect or recovered mask as a temporal reference.
1549
+ if not did_recover and np.count_nonzero(recovered >= 128) >= recovered.size * 0.015:
1550
  previous_alpha = recovered.copy()
1551
  blended = False
1552
  if active_session is not None:
 
1570
  results.append({"data":base64.b64encode(buf.getvalue()).decode("ascii"),
1571
  "format":fmt,"size":buf.tell(),
1572
  "model":used_model,"smoothed":blended,
1573
+ "matting":used_matting,"dropout_recovered":did_recover,
1574
+ "raw_foreground_fraction":raw_fraction,
1575
+ "final_foreground_fraction":final_fraction})
1576
  except Exception as e:
1577
  logger.error(f"[frames-batch] frame {idx} failed: {e}")
1578
  results.append({"data":"","format":"png","error":str(e)[:120]})