vikba commited on
Commit
496c9dc
·
verified ·
1 Parent(s): ae887fc

Upload folder using huggingface_hub

Browse files
PIPELINE.md CHANGED
@@ -5,7 +5,7 @@
5
  - Covers GUI and CLI startup, segmentation, measurement, and batch execution/export.
6
 
7
  ## Runtime Modes and Flags
8
- - `segmentation_mode=watershed` makes `analysis.segment(...)` run a gradient-masked watershed (`watershed_grad_threshold`).
9
  - `segmentation_mode=radial_expansion` makes `analysis.segment(...)` rasterize the expanded ridge fragments.
10
  - `merge_expansion_factor` controls overlap-based ridge-fragment merging (default `0.7`).
11
  - `segmentation_expansion_factor` controls radial label expansion (default `1.0`).
@@ -53,12 +53,12 @@
53
  - pads to the model stride, runs `OnnxFiberposeModel.predict(...)` -> `(dt, foreground_prob)`; `nout != 2` uses the last channel as DT and has no foreground
54
  - smooths DT, computes gradients/laplacian; stored via `Frame.set_prediction_outputs(...)`
55
  2. If the frame's segmentation signature (prediction signature + `segmentation_params(config)`) matches, stop.
56
- 3. Otherwise `clear_segmentation()` and run `analysis.segment(dt, dt_u8, gx, gy, foreground_prob, **segmentation_params)`:
57
  - `RidgeDetector.detect_ridges(...)` (raw ridge lines)
58
  - `_group_detected_polylines_into_segmentation_fragments(...)` merges ridges whose `merge_expansion_factor * DT` disks overlap
59
  - `_filter_measurement_centerlines(...)` derives measurement centerlines; the shortest fraction is dropped
60
  - `_filter_fragments_by_grad_quality(...)` drops fragments with unreliable DT gradients
61
- - watershed thresholds the DT gradient magnitude, unions two 3x3 dilations of the centerline markers, then applies erode -> dilate; radial_expansion rasterizes the expanded fragments
62
  - foreground, when available, prunes labels; only fragments that keep a label are returned
63
  4. Store the labels and fragments via `set_segmentation_outputs(...)`. The pipeline draws no images.
64
 
 
5
  - Covers GUI and CLI startup, segmentation, measurement, and batch execution/export.
6
 
7
  ## Runtime Modes and Flags
8
+ - `segmentation_mode=watershed` makes `analysis.segment(...)` run a watershed inside a fiber mask (`watershed_grad_threshold` sets its gradient part).
9
  - `segmentation_mode=radial_expansion` makes `analysis.segment(...)` rasterize the expanded ridge fragments.
10
  - `merge_expansion_factor` controls overlap-based ridge-fragment merging (default `0.7`).
11
  - `segmentation_expansion_factor` controls radial label expansion (default `1.0`).
 
53
  - pads to the model stride, runs `OnnxFiberposeModel.predict(...)` -> `(dt, foreground_prob)`; `nout != 2` uses the last channel as DT and has no foreground
54
  - smooths DT, computes gradients/laplacian; stored via `Frame.set_prediction_outputs(...)`
55
  2. If the frame's segmentation signature (prediction signature + `segmentation_params(config)`) matches, stop.
56
+ 3. Otherwise `clear_segmentation()` and run `analysis.segment(dt, dt_u8, gx, gy, foreground_prob, laplacian, **segmentation_params)`:
57
  - `RidgeDetector.detect_ridges(...)` (raw ridge lines)
58
  - `_group_detected_polylines_into_segmentation_fragments(...)` merges ridges whose `merge_expansion_factor * DT` disks overlap
59
  - `_filter_measurement_centerlines(...)` derives measurement centerlines; the shortest fraction is dropped
60
  - `_filter_fragments_by_grad_quality(...)` drops fragments with unreliable DT gradients
61
+ - watershed builds its mask from the DT gradient magnitude above the threshold or a negative Laplacian (ridge), keeps it only where the smoothed DT is positive, unions two 3x3 dilations of the centerline markers, then applies erode -> dilate; radial_expansion rasterizes the expanded fragments
62
  - foreground, when available, prunes labels; only fragments that keep a label are returned
63
  4. Store the labels and fragments via `set_segmentation_outputs(...)`. The pipeline draws no images.
64
 
README.md CHANGED
@@ -86,7 +86,7 @@ Running `pip install -e .` again reinstalls the CPU build, so repeat these two c
86
 
87
  ## Pretrained model
88
 
89
- The default model, `R108_dt_fg.onnx`, ships with the repository in `checkpoints/`, so a source checkout works offline. Fiberpose looks for models, by file name, in this order:
90
 
91
  1. the folder set in the `FIBERPOSE_MODELS_DIR` environment variable;
92
  2. `checkpoints/` in the repository, when running from a source checkout;
@@ -272,7 +272,7 @@ Fiberpose/
272
  ├── app.py # Entry point of the Hugging Face Space
273
  ├── requirements.txt # Dependencies of the Hugging Face Space
274
  ├── Dockerfile # Web interface as a container
275
- ├── checkpoints/ # Pretrained model (R108_dt_fg.onnx)
276
  ├── docs/ # README images
277
  └── src/fiberpose/
278
  ├── cli.py # GUI/CLI/web launcher (`fiberpose` command)
@@ -304,7 +304,7 @@ Fiberpose/
304
 
305
  ### Model file not found
306
 
307
- The default model is expected in `checkpoints/` of the source checkout. If Fiberpose runs from elsewhere, copy `R108_dt_fg.onnx` into a folder and point `FIBERPOSE_MODELS_DIR` at it. For a custom model, check that `model_path` in your configuration is a file name in the models folder or an absolute path.
308
 
309
  ### GPU is not used
310
 
 
86
 
87
  ## Pretrained model
88
 
89
+ The default model, `R108_dt_R111_fg.onnx`, ships with the repository in `checkpoints/`, so a source checkout works offline. Fiberpose looks for models, by file name, in this order:
90
 
91
  1. the folder set in the `FIBERPOSE_MODELS_DIR` environment variable;
92
  2. `checkpoints/` in the repository, when running from a source checkout;
 
272
  ├── app.py # Entry point of the Hugging Face Space
273
  ├── requirements.txt # Dependencies of the Hugging Face Space
274
  ├── Dockerfile # Web interface as a container
275
+ ├── checkpoints/ # Pretrained models (default: R108_dt_R111_fg.onnx)
276
  ├── docs/ # README images
277
  └── src/fiberpose/
278
  ├── cli.py # GUI/CLI/web launcher (`fiberpose` command)
 
304
 
305
  ### Model file not found
306
 
307
+ The default model is expected in `checkpoints/` of the source checkout. If Fiberpose runs from elsewhere, copy `R108_dt_R111_fg.onnx` into a folder and point `FIBERPOSE_MODELS_DIR` at it. For a custom model, check that `model_path` in your configuration is a file name in the models folder or an absolute path.
308
 
309
  ### GPU is not used
310
 
checkpoints/R108_dt_R111_fg.onnx ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8f89b9161cb454968d36fc7e6ea9c6d770a6dd45a891f0aa308d86e99778c6c0
3
+ size 5610733
src/fiberpose/analysis.py CHANGED
@@ -32,6 +32,12 @@ logger = logging.getLogger(__name__)
32
 
33
  Sample = tuple[int, int, float, float, float, float]
34
 
 
 
 
 
 
 
35
 
36
  # ------------------------------------------------------------------ #
37
  # Public pipeline stages
@@ -45,7 +51,7 @@ def predict(model: OnnxFiberposeModel, image: np.ndarray) -> dict[str, np.ndarra
45
  dt_pad, fg_pad = model.predict(np.pad(image, pad_width, mode="reflect"))
46
 
47
  dt = np.asarray(dt_pad[:h, :w], dtype=np.float32)
48
- dt_smooth = gaussian_filter(dt, sigma=1.5)
49
  gy, gx = np.gradient(dt_smooth)
50
  return {
51
  "dt": dt,
@@ -64,6 +70,7 @@ def segment(
64
  gx: np.ndarray,
65
  gy: np.ndarray,
66
  foreground_prob: np.ndarray | None,
 
67
  *,
68
  segmentation_mode: str,
69
  watershed_grad_threshold: float,
@@ -138,7 +145,7 @@ def segment(
138
  labels = np.clip(labels_i32, 0, np.iinfo(np.uint16).max).astype(np.uint16, copy=False)
139
  else:
140
  markers = _render_markers(shape, fragments)
141
- labels = _watershed(dt_u8, gx, gy, markers, gradient_threshold=watershed_grad_threshold)
142
  labels = _prune_labels_by_foreground(
143
  labels,
144
  foreground_prob=foreground_prob,
@@ -513,26 +520,35 @@ def _clamp_trim_fraction(value: float) -> float:
513
 
514
 
515
  def _watershed(
 
516
  dt_u8: np.ndarray,
517
  gx: np.ndarray,
518
  gy: np.ndarray,
 
519
  markers: np.ndarray,
520
  gradient_threshold: float,
521
  ) -> np.ndarray:
522
  """
523
- Watershed on inverted DT, constrained by thresholded gradient magnitude.
 
 
 
 
 
524
 
525
- Two dilations of the detected centerline markers are added to the
526
- threshold mask. The combined mask is then eroded and dilated before
527
- watershed.
528
  """
529
 
 
530
  dt_u8 = np.asarray(dt_u8)
531
  gx = np.asarray(gx, dtype=np.float32)
532
  gy = np.asarray(gy, dtype=np.float32)
 
533
  markers = np.asarray(markers)
534
- if gx.shape != dt_u8.shape or gy.shape != dt_u8.shape or markers.shape != dt_u8.shape:
535
- raise ValueError("DT, gradients, and markers must have matching shapes.")
536
  if not markers.any():
537
  return np.zeros_like(dt_u8, dtype=np.uint16)
538
 
@@ -543,6 +559,8 @@ def _watershed(
543
  dt_norm = dt_u8.astype(np.float32) / 255.0
544
  grad_magnitude = np.hypot(gx, gy)
545
  fiber_region_mask = np.isfinite(grad_magnitude) & (grad_magnitude >= threshold)
 
 
546
 
547
  morphology_kernel = np.ones((3, 3), dtype=np.uint8)
548
  centerline_mask = (markers > 0).astype(np.uint8)
@@ -569,7 +587,7 @@ def _watershed(
569
  return np.zeros_like(dt_u8, dtype=np.uint16)
570
  marker_pixels = markers > 0
571
  logger.info(
572
- "Watershed gradient mask: threshold=%.4f mask=%d/%d marker_pixels=%d/%d",
573
  threshold,
574
  int(np.count_nonzero(fiber_region_mask)),
575
  int(fiber_region_mask.size),
 
32
 
33
  Sample = tuple[int, int, float, float, float, float]
34
 
35
+ # Smoothing applied to the predicted DT before differentiating it; the watershed
36
+ # mask gates on the same smoothed DT, so both must use this value.
37
+ _DT_SMOOTH_SIGMA = 1.5
38
+ # A fiber centerline is a ridge of the DT: its Laplacian is clearly negative.
39
+ _RIDGE_LAPLACIAN_THRESHOLD = 0.03
40
+
41
 
42
  # ------------------------------------------------------------------ #
43
  # Public pipeline stages
 
51
  dt_pad, fg_pad = model.predict(np.pad(image, pad_width, mode="reflect"))
52
 
53
  dt = np.asarray(dt_pad[:h, :w], dtype=np.float32)
54
+ dt_smooth = gaussian_filter(dt, sigma=_DT_SMOOTH_SIGMA)
55
  gy, gx = np.gradient(dt_smooth)
56
  return {
57
  "dt": dt,
 
70
  gx: np.ndarray,
71
  gy: np.ndarray,
72
  foreground_prob: np.ndarray | None,
73
+ laplacian: np.ndarray,
74
  *,
75
  segmentation_mode: str,
76
  watershed_grad_threshold: float,
 
145
  labels = np.clip(labels_i32, 0, np.iinfo(np.uint16).max).astype(np.uint16, copy=False)
146
  else:
147
  markers = _render_markers(shape, fragments)
148
+ labels = _watershed(dt, dt_u8, gx, gy, laplacian, markers, gradient_threshold=watershed_grad_threshold)
149
  labels = _prune_labels_by_foreground(
150
  labels,
151
  foreground_prob=foreground_prob,
 
520
 
521
 
522
  def _watershed(
523
+ dt: np.ndarray,
524
  dt_u8: np.ndarray,
525
  gx: np.ndarray,
526
  gy: np.ndarray,
527
+ laplacian: np.ndarray,
528
  markers: np.ndarray,
529
  gradient_threshold: float,
530
  ) -> np.ndarray:
531
  """
532
+ Watershed on inverted DT, constrained to a fiber mask.
533
+
534
+ The mask is the fiber flanks (gradient magnitude above the threshold) plus
535
+ the ridges between them (negative Laplacian, where the gradient vanishes),
536
+ both limited to where the smoothed DT is positive: the predicted background
537
+ is negative and sloped, so neither test alone tells it from a fiber.
538
 
539
+ Two dilations of the detected centerline markers are added to the mask,
540
+ which keeps fibers too thin for the smoothed DT to stay positive. The
541
+ combined mask is then eroded and dilated before watershed.
542
  """
543
 
544
+ dt = np.asarray(dt, dtype=np.float32)
545
  dt_u8 = np.asarray(dt_u8)
546
  gx = np.asarray(gx, dtype=np.float32)
547
  gy = np.asarray(gy, dtype=np.float32)
548
+ laplacian = np.asarray(laplacian, dtype=np.float32)
549
  markers = np.asarray(markers)
550
+ if any(arr.shape != dt_u8.shape for arr in (dt, gx, gy, laplacian, markers)):
551
+ raise ValueError("DT, gradients, Laplacian, and markers must have matching shapes.")
552
  if not markers.any():
553
  return np.zeros_like(dt_u8, dtype=np.uint16)
554
 
 
559
  dt_norm = dt_u8.astype(np.float32) / 255.0
560
  grad_magnitude = np.hypot(gx, gy)
561
  fiber_region_mask = np.isfinite(grad_magnitude) & (grad_magnitude >= threshold)
562
+ fiber_region_mask |= laplacian < -_RIDGE_LAPLACIAN_THRESHOLD
563
+ fiber_region_mask &= gaussian_filter(dt, sigma=_DT_SMOOTH_SIGMA) > 0.0
564
 
565
  morphology_kernel = np.ones((3, 3), dtype=np.uint8)
566
  centerline_mask = (markers > 0).astype(np.uint8)
 
587
  return np.zeros_like(dt_u8, dtype=np.uint16)
588
  marker_pixels = markers > 0
589
  logger.info(
590
+ "Watershed fiber mask: threshold=%.4f mask=%d/%d marker_pixels=%d/%d",
591
  threshold,
592
  int(np.count_nonzero(fiber_region_mask)),
593
  int(fiber_region_mask.size),
src/fiberpose/models.py CHANGED
@@ -16,7 +16,7 @@ from pathlib import Path
16
 
17
  import pooch
18
 
19
- DEFAULT_MODEL = "R108_dt_fg.onnx"
20
 
21
  # TODO: upload the registered files as assets of this release before publishing.
22
  _BASE_URL = "https://github.com/vikba/Fiberpose/releases/download/models-v1/"
@@ -24,10 +24,12 @@ _BASE_URL = "https://github.com/vikba/Fiberpose/releases/download/models-v1/"
24
  # Model file -> files to download; external-data files (.onnx.data) must sit
25
  # next to their .onnx graph.
26
  _MODELS: dict[str, tuple[str, ...]] = {
 
27
  "R108_dt_fg.onnx": ("R108_dt_fg.onnx",),
28
  }
29
 
30
  _REGISTRY = {
 
31
  "R108_dt_fg.onnx": "sha256:bb515be44e1daa220f8ba457fb1f9d856a2c3865149f4b6b97e12aefdea37536",
32
  }
33
 
 
16
 
17
  import pooch
18
 
19
+ DEFAULT_MODEL = "R108_dt_R111_fg.onnx"
20
 
21
  # TODO: upload the registered files as assets of this release before publishing.
22
  _BASE_URL = "https://github.com/vikba/Fiberpose/releases/download/models-v1/"
 
24
  # Model file -> files to download; external-data files (.onnx.data) must sit
25
  # next to their .onnx graph.
26
  _MODELS: dict[str, tuple[str, ...]] = {
27
+ "R108_dt_R111_fg.onnx": ("R108_dt_R111_fg.onnx",),
28
  "R108_dt_fg.onnx": ("R108_dt_fg.onnx",),
29
  }
30
 
31
  _REGISTRY = {
32
+ "R108_dt_R111_fg.onnx": "sha256:8f89b9161cb454968d36fc7e6ea9c6d770a6dd45a891f0aa308d86e99778c6c0",
33
  "R108_dt_fg.onnx": "sha256:bb515be44e1daa220f8ba457fb1f9d856a2c3865149f4b6b97e12aefdea37536",
34
  }
35
 
src/fiberpose/pipeline.py CHANGED
@@ -335,7 +335,13 @@ class PipelineController:
335
  frame.clear_segmentation()
336
  try:
337
  labels, fragments = analysis.segment(
338
- frame.dt, frame.dt_u8, frame.gx, frame.gy, frame.foreground_prob, **self.segmentation_params(config)
 
 
 
 
 
 
339
  )
340
  except Exception as exc:
341
  frame.fail(str(exc))
 
335
  frame.clear_segmentation()
336
  try:
337
  labels, fragments = analysis.segment(
338
+ frame.dt,
339
+ frame.dt_u8,
340
+ frame.gx,
341
+ frame.gy,
342
+ frame.foreground_prob,
343
+ frame.lapl_img,
344
+ **self.segmentation_params(config),
345
  )
346
  except Exception as exc:
347
  frame.fail(str(exc))
src/fiberpose/resources/config.json CHANGED
@@ -1,5 +1,5 @@
1
  {
2
- "model_path": "R108_dt_fg.onnx",
3
  "use_gpu": true,
4
  "scale_label": "1x",
5
  "max_angle_deg": 30.0,
 
1
  {
2
+ "model_path": "R108_dt_R111_fg.onnx",
3
  "use_gpu": true,
4
  "scale_label": "1x",
5
  "max_angle_deg": 30.0,
src/fiberpose/resources/fiberpose.png CHANGED
src/fiberpose/web/demo.py CHANGED
@@ -59,7 +59,9 @@ def build_demo(config: AppConfig | None = None) -> gr.Blocks:
59
  "# Fiberpose\nSegmentation and morphometry of fibrous microstructures. Upload a micrograph "
60
  f"up to {cb.MAX_SIDE} x {cb.MAX_SIDE} px (or click an example) and press **Analyze**. [Source code](https://github.com/vikba/Fiberpose)"
61
  "\n\nIf you use Fiberpose in scientific work, please cite the preprint: "
62
- "[arXiv:2611.22953](https://doi.org/10.48550/arXiv.2611.22953)."
 
 
63
  )
64
  session = gr.State() # this browser tab's analysed image (callbacks.Session)
65
  with gr.Row(equal_height=True): # settings, each above the image it applies to
 
59
  "# Fiberpose\nSegmentation and morphometry of fibrous microstructures. Upload a micrograph "
60
  f"up to {cb.MAX_SIDE} x {cb.MAX_SIDE} px (or click an example) and press **Analyze**. [Source code](https://github.com/vikba/Fiberpose)"
61
  "\n\nIf you use Fiberpose in scientific work, please cite the preprint: "
62
+ "V. A. Balashov, D. S. Antonenko. *Fiberpose: state-of-the-art fiber morphometry and instance "
63
+ "segmentation with a distance-transform neural network.* "
64
+ "[arXiv:2611.22953](https://doi.org/10.48550/arXiv.2611.22953) (2026)."
65
  )
66
  session = gr.State() # this browser tab's analysed image (callbacks.Session)
67
  with gr.Row(equal_height=True): # settings, each above the image it applies to