ROIを頭部だけに絞ってアップスケールする

config/default.yaml

@@ -72,8 +72,10 @@ face:
   max_persons_per_frame: 3
   require_confirmed_track: true
   process_only_inside_region: false
-  roi_height_ratio: 0.60
+  roi_height_ratio: 0.25
   roi_expand_ratio: 0.20
+  roi_square: true
+  roi_upscale_target: 384
   min_roi_width: 40
   min_roi_height: 40
   static_image_mode: false

src/config_loader.py

@@ -181,6 +181,8 @@ class FaceSettings:
     process_only_inside_region: bool
     roi_height_ratio: float
     roi_expand_ratio: float
+    roi_square: bool
+    roi_upscale_target: int
     min_roi_width: int
     min_roi_height: int
     static_image_mode: bool
@@ -574,6 +576,8 @@ def _validate(config: AppConfig) -> None:
         raise ConfigurationError("face ROI ratios are outside their valid range")
     if face.min_roi_width <= 0 or face.min_roi_height <= 0:
         raise ConfigurationError("face minimum ROI dimensions must be positive")
+    if face.roi_upscale_target < 0 or not isinstance(face.roi_square, bool):
+        raise ConfigurationError("face roi_square/roi_upscale_target is invalid")
     if not 0.0 <= face.min_detection_confidence <= 1.0 or not 0.0 <= face.min_tracking_confidence <= 1.0:
         raise ConfigurationError("face confidence thresholds must be in [0, 1]")
     pose = config.head_pose

src/face_geometry.py

@@ -14,10 +14,15 @@ def create_face_roi(person: TrackedPerson, frame_width: int, frame_height: int,
     candidate_height = bbox.height * config.roi_height_ratio
     expand_x = bbox.width * config.roi_expand_ratio
     expand_y = candidate_height * config.roi_expand_ratio
-    x1 = max(0, round(bbox.x1 - expand_x))
     y1 = max(0, round(bbox.y1 - expand_y))
-    x2 = min(frame_width, round(bbox.x2 + expand_x))
     y2 = min(frame_height, round(bbox.y1 + candidate_height + expand_y))
+    if config.roi_square:
+        half = (y2 - y1) / 2.0          # 縦幅と同じ幅の正方形
+        x1 = max(0, round(bbox.center_x - half))
+        x2 = min(frame_width, round(bbox.center_x + half))
+    else:
+        x1 = max(0, round(bbox.x1 - expand_x))
+        x2 = min(frame_width, round(bbox.x2 + expand_x))
     if x2 - x1 < config.min_roi_width or y2 - y1 < config.min_roi_height:
         return None
     return FaceROI(person.track_id, x1, y1, x2, y2, bbox, person.confidence)

src/face_processor.py

@@ -71,6 +71,11 @@ class FaceProcessor:
             image = frame[roi.y1:roi.y2, roi.x1:roi.x2]
             import cv2
 
+            target = self.config.roi_upscale_target
+            longest = max(image.shape[:2])
+            if target > 0 and longest < target:
+                scale = target / longest
+                image = cv2.resize(image, None, fx=scale, fy=scale, interpolation=cv2.INTER_CUBIC)
             output = self._mesh.process(cv2.cvtColor(image, cv2.COLOR_BGR2RGB))
             faces = getattr(output, "multi_face_landmarks", None) or []
             if not faces: