Update landmarkdiff/face_verifier.py to v0.3.2

landmarkdiff/face_verifier.py

@@ -15,6 +15,7 @@ Designed for:
 
 from __future__ import annotations
 
+import logging
 from dataclasses import dataclass, field
 from pathlib import Path
 from typing import Any
@@ -22,6 +23,8 @@ from typing import Any
 import cv2
 import numpy as np
 
+logger = logging.getLogger(__name__)
+
 # ---------------------------------------------------------------------------
 # Data structures
 # ---------------------------------------------------------------------------
@@ -706,7 +709,11 @@ def get_face_embedding(image: np.ndarray) -> np.ndarray | None:
     try:
         faces = app.get(image)
         if faces:
-            return faces[0].embedding
+            emb = faces[0].embedding
+            if np.linalg.norm(emb) < 1e-6:
+                logger.warning("ArcFace returned near-zero embedding (occluded face?)")
+                return None
+            return emb
     except Exception:
         pass
     return None
@@ -715,12 +722,12 @@ def get_face_embedding(image: np.ndarray) -> np.ndarray | None:
 def verify_identity(
     original: np.ndarray,
     restored: np.ndarray,
-    threshold: float = 0.6,
+    threshold: float = 0.5,
 ) -> tuple[float, bool]:
     """Compare identity between original and restored using ArcFace.
 
     Returns (cosine_similarity, passed).
-    Similarity > threshold means same person (threshold=0.6 is conservative).
+    Similarity > threshold means same person (0.5 accommodates non-frontal poses).
     """
     emb_orig = get_face_embedding(original)
     emb_rest = get_face_embedding(restored)
@@ -743,11 +750,11 @@ def verify_identity(
 def verify_and_restore(
     image: np.ndarray,
     quality_threshold: float = 60.0,
-    identity_threshold: float = 0.6,
+    identity_threshold: float = 0.5,
     restore_mode: str = "auto",
     codeformer_fidelity: float = 0.7,
 ) -> RestorationResult:
-    """Full pipeline: analyze → restore → verify identity.
+    """Full pipeline: analyze -> restore -> verify identity.
 
     This is the main entry point for the face verifier. It:
     1. Analyzes the input for distortions
@@ -830,7 +837,7 @@ def verify_batch(
     image_dir: str,
     output_dir: str | None = None,
     quality_threshold: float = 60.0,
-    identity_threshold: float = 0.6,
+    identity_threshold: float = 0.5,
     restore_mode: str = "auto",
     save_rejected: bool = False,
     extensions: tuple[str, ...] = (".jpg", ".jpeg", ".png", ".webp", ".bmp"),
@@ -882,7 +889,7 @@ def verify_batch(
 
     for i, img_file in enumerate(image_files):
         if (i + 1) % 50 == 0 or i == 0:
-            print(f"Processing {i + 1}/{len(image_files)}: {img_file.name}")
+            logger.info("Processing %d/%d: %s", i + 1, len(image_files), img_file.name)
 
         image = cv2.imread(str(img_file))
         if image is None:
@@ -934,7 +941,7 @@ def verify_batch(
     # Save report
     report_text = report.summary()
     (out_path / "report.txt").write_text(report_text)
-    print(f"\n{report_text}")
-    print(f"\nResults saved to {out_path}/")
+    logger.info("\n%s", report_text)
+    logger.info("Results saved to %s/", out_path)
 
     return report