Switch to OpenCV face detection

app.py

@@ -9,7 +9,6 @@ import time
 import cv2
 import numpy as np
 import onnxruntime as ort
-import mediapipe as mp
 import gradio as gr
 from PIL import Image
 from huggingface_hub import hf_hub_download
@@ -37,41 +36,75 @@ sess = ort.InferenceSession(
 print("Model loaded.")
 
 # ---------------------------------------------------------------------------
-# MediaPipe face detection
+# OpenCV DNN face detection (no extra dependencies)
 # ---------------------------------------------------------------------------
-mp_face_detection = mp.solutions.face_detection
+# Use OpenCV's built-in Haar cascade as primary, with DNN SSD as fallback
+_face_cascade = cv2.CascadeClassifier(
+    cv2.data.haarcascades + "haarcascade_frontalface_default.xml"
+)
+
+# Try to use the more accurate DNN face detector if available
+_dnn_net = None
+_dnn_model_path = os.path.join(os.path.dirname(__file__), "face_detection_yunet_2023mar.onnx")
+YUNET_URL = "https://github.com/opencv/opencv_zoo/raw/main/models/face_detection_yunet/face_detection_yunet_2023mar.onnx"
+
+def _ensure_yunet():
+    """Download YuNet face detector if not present."""
+    global _dnn_model_path
+    if not os.path.exists(_dnn_model_path):
+        print("Downloading YuNet face detector...")
+        try:
+            path = hf_hub_download(
+                repo_id="opencv/opencv_zoo",
+                filename="models/face_detection_yunet/face_detection_yunet_2023mar.onnx",
+            )
+            _dnn_model_path = path
+        except Exception:
+            import urllib.request
+            urllib.request.urlretrieve(YUNET_URL, _dnn_model_path)
+        print("YuNet downloaded.")
+    return _dnn_model_path
+
 
 def detect_face_box(image_rgb: np.ndarray):
     """
-    Detect the largest face bounding box using MediaPipe.
+    Detect the largest face bounding box.
     Returns (x1, y1, x2, y2) in pixel coords or None.
     """
     h, w = image_rgb.shape[:2]
-    with mp_face_detection.FaceDetection(
-        model_selection=1, min_detection_confidence=0.5
-    ) as face_det:
-        results = face_det.process(image_rgb)
-        if not results.detections:
-            return None
-
-        # Pick the largest detection by area
-        best = None
-        best_area = 0
-        for det in results.detections:
-            bb = det.location_data.relative_bounding_box
-            area = bb.width * bb.height
-            if area > best_area:
-                best_area = area
-                best = bb
-
-        # Convert relative to absolute
-        x1 = int(best.xmin * w)
-        y1 = int(best.ymin * h)
-        bw = int(best.width * w)
-        bh = int(best.height * h)
-        x2 = x1 + bw
-        y2 = y1 + bh
-        return (x1, y1, x2, y2)
+
+    # Try YuNet first (more accurate)
+    try:
+        yunet_path = _ensure_yunet()
+        detector = cv2.FaceDetectorYN.create(yunet_path, "", (w, h), 0.5, 0.3, 5000)
+        _, faces = detector.detect(image_rgb)
+        if faces is not None and len(faces) > 0:
+            # Pick largest face by area
+            best_idx = 0
+            best_area = 0
+            for i, face in enumerate(faces):
+                fw, fh = face[2], face[3]
+                area = fw * fh
+                if area > best_area:
+                    best_area = area
+                    best_idx = i
+            f = faces[best_idx]
+            x1, y1 = int(f[0]), int(f[1])
+            x2, y2 = int(f[0] + f[2]), int(f[1] + f[3])
+            return (max(x1, 0), max(y1, 0), min(x2, w), min(y2, h))
+    except Exception as e:
+        print(f"YuNet failed, falling back to Haar: {e}")
+
+    # Fallback: Haar cascade
+    gray = cv2.cvtColor(image_rgb, cv2.COLOR_RGB2GRAY)
+    faces = _face_cascade.detectMultiScale(gray, scaleFactor=1.1, minNeighbors=5, minSize=(60, 60))
+    if len(faces) == 0:
+        return None
+
+    # Pick largest
+    best_idx = np.argmax([fw * fh for (_, _, fw, fh) in faces])
+    x, y, fw, fh = faces[best_idx]
+    return (x, y, x + fw, y + fh)
 
 # ---------------------------------------------------------------------------
 # Face cropping with margin

requirements.txt

@@ -3,4 +3,3 @@ onnxruntime
 opencv-python-headless
 numpy
 Pillow
-mediapipe==0.10.14