Update landmarkdiff/landmarks.py to v0.3.2

landmarkdiff/landmarks.py

@@ -2,6 +2,7 @@
 
 from __future__ import annotations
 
+import logging
 from dataclasses import dataclass
 from pathlib import Path
 
@@ -9,6 +10,8 @@ import cv2
 import mediapipe as mp
 import numpy as np
 
+logger = logging.getLogger(__name__)
+
 # Region color map for visualization (BGR)
 REGION_COLORS: dict[str, tuple[int, int, int]] = {
     "jawline": (255, 255, 255),  # white
@@ -167,12 +170,45 @@ class FaceLandmarks:
 
     @property
     def pixel_coords(self) -> np.ndarray:
-        """Convert normalized landmarks to pixel coordinates (478, 2)."""
+        """Convert normalized landmarks to pixel coordinates (478, 2).
+
+        Coordinates are clamped to valid image bounds so that extreme
+        head poses do not produce out-of-range indices.
+        """
         coords = self.landmarks[:, :2].copy()
         coords[:, 0] *= self.image_width
         coords[:, 1] *= self.image_height
+        coords[:, 0] = np.clip(coords[:, 0], 0, self.image_width - 1)
+        coords[:, 1] = np.clip(coords[:, 1], 0, self.image_height - 1)
+        return coords
+
+    def pixel_coords_at(self, width: int, height: int) -> np.ndarray:
+        """Convert normalized landmarks to pixel coordinates at a given size.
+
+        Use this when the image has been resized after landmark extraction.
+        Coordinates are clamped to [0, width-1] x [0, height-1].
+        """
+        coords = self.landmarks[:, :2].copy()
+        coords[:, 0] *= width
+        coords[:, 1] *= height
+        coords[:, 0] = np.clip(coords[:, 0], 0, width - 1)
+        coords[:, 1] = np.clip(coords[:, 1], 0, height - 1)
         return coords
 
+    def rescale(self, width: int, height: int) -> FaceLandmarks:
+        """Return a copy with updated image dimensions.
+
+        Landmarks stay in normalized [0,1] space; only the stored
+        width/height change, so ``pixel_coords`` returns values at
+        the new resolution.
+        """
+        return FaceLandmarks(
+            landmarks=self.landmarks.copy(),
+            image_width=width,
+            image_height=height,
+            confidence=self.confidence,
+        )
+
     def get_region(self, region: str) -> np.ndarray:
         """Get landmark indices for a named region."""
         indices = LANDMARK_REGIONS.get(region, [])
@@ -201,11 +237,13 @@ def extract_landmarks(
     try:
         landmarks, confidence = _extract_tasks_api(rgb, min_detection_confidence)
     except Exception:
+        logger.debug("Tasks API unavailable, trying Solutions API", exc_info=True)
         try:
             landmarks, confidence = _extract_solutions_api(
                 rgb, min_detection_confidence, min_tracking_confidence
             )
         except Exception:
+            logger.debug("Both MediaPipe APIs failed", exc_info=True)
             return None
 
     if landmarks is None:
@@ -338,7 +376,7 @@ def render_landmark_image(
     """Render MediaPipe face mesh tessellation on black canvas.
 
     Draws the full 2556-edge tessellation mesh that CrucibleAI/ControlNetMediaPipeFace
-    was pre-trained on. This is critical — the ControlNet expects dense triangulated
+    was pre-trained on. This is critical -- the ControlNet expects dense triangulated
     wireframes, not sparse dots.
 
     Falls back to colored dots if tessellation connections aren't available.
@@ -380,7 +418,7 @@ def render_landmark_image(
             p2 = tuple(pts[conn.end])
             cv2.line(canvas, p1, p2, (255, 255, 255), 1, cv2.LINE_AA)
 
-    except ImportError:
+    except (ImportError, AttributeError):
         # Fallback: draw colored dots if tessellation not available
         idx_to_color: dict[int, tuple[int, int, int]] = {}
         for region, indices in LANDMARK_REGIONS.items():